Dispute analysis within the team: does optimization for generative engines exist as a separate discipline, and what does this imply for kithrabbit.com and for community pages on their own domains, such as embodied.nexus. The main question is posed directly: is SEO enough for these pages, or do they need GEO and AEO.
The dispute is short. One side sends a table where optimization for generative engines looks like a new discipline with different goals, different signals, and different measurements. The second side replies that it's the same SEO and the same SMM, and doesn't require separate work. Both sides are partially right, and this can only be verified with numbers, which simply didn't exist until 2026.
The numbers are as follows. A direct log experiment on a high-traffic domain showed that the raw growth of transitions from ChatGPT was ×5.7, but unprocessed pages of the same domain grew ×3.5 during the same window. That is, most of the success commonly attributed to optimization is a tailwind from the platform itself. The estimation of the actual causal effect yielded ×1.82 (95% confidence interval 1.31–2.54, p=0.001), but a placebo-in-time permutation test gave p=0.16, meaning the effect is indicative, not proven. The authors explicitly state: the headline multiples of AEO significantly overestimate the causal effect.
And yet, three differences from SEO are based on measured data, not promises. The unit is no longer a position, but an inclusion in the answer text. Measurement ceases to be deterministic: identical questions yield different answers and different sources. And a step appears that SEO never had: the agent visits the site and reads it before making a decision.
And immediately, the answer to the question for which everything is gathered. For kithrabbit.com and for community pages on their own domains, such as embodied.nexus, SEO is not enough, but what is usually sold as GEO is also not enough. The root cause is not in the texts or in mentions: a measurement of nine network sites showed that no page publishes structural markup, only one provides a description for the agent, only two have crawl rules and a sitemap, and none contain a verification statement. The embodied.nexus page is 236 words, zero numbers, zero dates, zero rules. There is nothing for the agent to read here, and this is a separate job that neither SEO nor community marketing handles.
The group discusses a six-row table, where on the left is classic SEO, and on the right is optimization for generative engines, named by two names: AISO, optimization for agent search, and GEO, optimization for generative engines. The table states that the goal changes, the end result for the user changes, ranking signals change, content format changes, the nature of competition changes, and the method of measurement changes.
Two arguments are made against the table. The first, from participant @james: if you make a good website and manage social networks, especially Reddit, then optimization will happen by itself, there's no need to do it separately. The second, from the organizer: he sees no reason to do anything other than what is done for SEO and SMM, and asks for research to be conducted and, if possible, to prove him wrong.
A separate discipline emerges: a different goal, different signals, different format, different measurement, all-or-nothing competition.
What's done for SEO and SMM is sufficient. There's no separate work, because the result is the same and the levers are the same.
The dispute cannot be closed with words, because both sides speak plausibly. It can only be closed with measurements, and these appeared in 2026 in quantities that were not present in 2023, when the first work on the topic was published, nor in 2024, when almost no one had heard of it.
Below, each row of the table is checked individually, then four decisive studies are analyzed, then industry measurements from vendors, then a conclusion for the Kith Rabbit network pages: kithrabbit.com itself and community pages on their own domains. Where the source is vendor-specific, this is explicitly stated, because vendor measurements do not pass scientific scrutiny and often measure their own interest.
The analysis addresses a specific dispute, so its positions should be fixed before the numbers. Five assertions were made in the discussion, and each is then checked separately.
| Assertion | Who | Where it is checked |
|---|---|---|
| Good structural content always wins, so it's enough to create more content, and the data itself will provide both SEO and GEO | organizer and participant | sections on multiples, on volume, and on readability |
| Models do not feel emotions, so people need emotional delivery, and machines need facts | participant | section on emotion as two channels |
| Models imitate emotions just like humans, so emotional content also works on them | organizer | the same section |
| Community pages should feed the Kith Rabbit page, including emotional attachment and a recognizable brand mark | participant | section on solutions for community pages |
| The community page belongs to the community; a small mark is enough for us, plus a verification badge for conversion | organizer | sections on the badge and on solutions |
Separately, two background assumptions, which are also being tested, flashed in the conversation: that the presence of communities on Kith Rabbit in itself means visibility of communities to agents, and that the verification badge works the same for a human and for a machine. Both assumptions were not confirmed during measurement, and below it is clear why.
The table can be checked line by line. Below, each line receives an assessment: confirmed, partially confirmed, or refuted. The assessment is made not by opinion, but by the presence of a measurement.
| Table row | Assertion | What the data shows | Assessment |
|---|---|---|---|
| Goal | Previously, position in search results, now mention within the answer | The unit of account is indeed different: there is no position in the answer, there is inclusion in the text and citation. The first source takes 40–55% of subsequent transitions, the distribution is log-normal. | Confirmed |
| Result for the user | Previously a list of pages, now one synthesized answer | This is a property of the product itself, not optimization, and it is true by definition: the generative engine returns text, not a list of links. | Confirmed |
| Ranking signals | Previously keywords, meta tags, and links; now meaning, entities, accuracy, and markup | Only partially confirmed. Link volume is indeed not the main factor, but what turns out to be main is not markup itself, but the density of extractable evidence: definitions, numbers, comparisons, step-by-step procedures. The specified author increases citability by 2.4 times. | Partially |
| Content format | Previously, pages for quick scanning; now, machine-readable structures and direct answers | The structure indeed works: merely restructuring without changing the meaning yields +17.3% in citations. But direct answers in a Q&A format do not, by themselves, improve comprehension; this is a proven negative result. | Partially |
| Competition | Previously, positions from first to tenth; now, "all or nothing" | The phrasing is too harsh. An answer often contains several brands: an average of 6.2 in one measurement and 3.7 in another. But the concentration is indeed high: the top-1 brand in the sector accounts for 31.4% of mentions, and the top-3 together account for 64.7%. It's more accurate to say "in or out of the answer, and the order within the answer is not guaranteed." | Partially |
| Measurement | Previously, clicks, impressions, positions; now, brand mentions, citations, conversation reach | This is confirmed, and it is the strongest distinction. Measurement becomes non-deterministic: identical questions yield different sources, so classical position-based attribution is impossible, and repeated runs and confidence intervals are needed. | Confirmed |
Table summary: two rows are fully confirmed, three partially, and one requires rephrasing. This is not "the same SEO" because the measurement and unit of account are indeed different. But neither are there six new laws, because half of the rows merely rephrase ordinary content work.
The main difference that the table correctly identified is measurement. While SEO is measured by position and click, optimization for engines is measured by inclusion in the text, and such a measurement is non-deterministic by nature. This is why the dispute cannot be resolved with old tools: they lack a counter for what is happening here.
The field is three years old, and its age explains why the debate is still alive. The first work where the term "generative engine optimization" appeared as a scientific concept was published in November 2023. It showed that content modification could increase citation frequency by up to forty percent. The work was published at a time when generative engines were just beginning to answer people's questions, and its conclusions were barely scrutinized by anyone.
Then came a pause. In 2024, engines primarily answered from training data and hallucinated when that data was insufficient, so advice boiled down to "seeding" the model's knowledge. This was a bad time for the topic: it looked like manipulation of training, not like website work.
The turning point occurred when models transitioned to live web search. Then, knowledge from training ceased to be primary: a measurement across almost forty thousand agent routes showed that only 7 to 10 percentof the training data makes it into the final answer. Everything else comes from live web reading, and this made working with a website meaningful again, but under different rules.
First work on the topic. The term appears in science, claiming up to +40% in citation rate from content changes.
Pause. Models respond from training, the myth of “knowledge seeding” appears. No systematic measurements.
Models move to live search. Working with a website is important again, but not as optimization for keywords.
Measurable frameworks emerge: two stages of citation, non-determinism, structural contribution, log-experiments with control.
The concept of agent readability emerges as a separate lever: thirty-seven thousand routes versus over a thousand businesses.
The debate ongoing in the group today is a debate within a field that is three years old and has one full measurement cycle behind it. This is why it can be answered with numbers, not with a position.
The most honest study of the field was conducted on a single domain with hundreds of thousands of pages. The owner applied a set of measures to one part of the site to optimize it for answer engines, while leaving the rest of the same site as is. This proved to be a strong move: the unprocessed pages of the same domain automatically received exactly the same platform acceleration as the processed ones, because ChatGPT grew rapidly on its own during that period.
The measurement yielded four numbers. Total traffic from ChatGPT increased by 5.7 times. Unprocessed pages of the same domain increased by 3.5 times. The difference between these two numbers is what can be attributed to the measures. The model, based on weekly data for processed versus control, showed a jump ×1.82 with an interval from 1.31 to 2.54 and a significance level of p=0.001, and with an engagement filter — ×2.27. But the placebo-in-time permutation test, which checks whether the same result could have occurred by chance, yielded p=0.16. This means: the effect is indicative, but not proven.
The study's conclusion is more important than any single number: the result of measures must be separated from platform growth by controlling on the same domain, and headline multiples in vendor reports significantly exaggerate the causal effect.
If you remove the tailwind, the effect remains around two times, and it is difficult to prove in a short window. This does not mean that working is pointless. It means that measurement should be done differently, and that most public success stories measure not the effect of measures, but the growth of the engines themselves.
The mechanics of deception are simple and repeated by everyone. The platform grows by double-digit percentages per quarter. The owner does something on the site. Transitions grow. Everything else on the site grows too, but the owner only looks at the changed pages. This is where multiples of five, seven, and ten times come from, which cannot be reproduced on another domain, because there is a different proportion of tailwind there.
Any report like “we increased citations sevenfold” almost certainly measures platform growth plus the effect of measures and attributes everything to the second term. The first question to such a report is simple: how much did the control grow, i.e., the unprocessed part of the same site? If there is no answer, the number means nothing.
And here the dispute gets the first half of the answer. The organizer is right that the loud promises of optimization specialists for engines are not confirmed: a significant part of the stated growth occurs without their participation. Critics of the table are right in the main point, and the table is only right in the description, not in the scale.
The most useful practical result of 2026 is that citation breaks down into two independent stages, and they must be calculated separately. The first stage is source selection: the engine starts a search and decides which pages to consider at all. The second stage is assimilationthe selected page actually gets into the answer text, bringing a formulation, number, comparison, or fact.
The separation of stages explains why the citation counter is misleading. Engines behave differently: one takes more sources and uses them less, another takes fewer and integrates them more deeply. Therefore, data on who cites more says nothing about who has a stronger influence on the answer.
The second result of the assimilation stage is even more important: pages with high impact are longer, more modular, and closer in meaning to the answer itself and contain extractable genres of evidence. Definitions, numerical facts, comparisons, and step-by-step procedures. That is, the engine does not look for text about you, but for ready-made bricks from which it will assemble an answer, and prefers pages where these bricks are already cut out and signed.
Optimization should not be about "being quoted," but about "what is chosen being useful to embed." These are different tasks: the first is about recognition and authority, the second is about the form and density of evidence on the page.
The same study has a conclusion that directly contradicts the most popular recommendation in this area. The "question-answer" format itself does not improve comprehension. Not at all. The presence of Q&A blocks does not increase the likelihood that the phrasing from the page will appear in the final answer.
This is more important than it looks. Most of the advice given to businesses boils down to adding a FAQ block to the page and markup for it. The measurement says that this is not enough. It's not the genre of the block that works, but the presence of extractable content within it: a precise definition, a specific number, an honest comparison, a reproducible sequence of actions.
Q&A block. Markup for the sake of markup. Direct one-line answers without content. Repetition of the key phrase.
Definitions, numbers, comparisons, step-by-step procedures. A specified author with a verifiable identity. Structure separating blocks from each other. Density of external primary sources.
Hence follows an unpleasant conclusion for the market: a significant part of the services sold in this area sells form, whereas the measured value lies in content. This is the second part of the answer to the organizer, and it is on his side: advice like "make a question block" is based on an unverified assumption.
A separate work separated structure from meaning and measured the contribution of structure in its pure form. The authors divided the page structure into three levels. The macro level is the document architecture: how it is broken down into sections and in what order. The meso level is the cutting into information blocks within sections. The micro level is visual accents, that is, how the reader's and parser's attention is directed within a block.
Measurement result: solely due to structure, with unchanged meaning, citation increased by 17.3 percent across six engines, and the subjective assessment of page quality increased by 18.5 percent. The authors describe this as moving work from the realm of intuition to the realm of engineering rules.
What this means in practice. Structure is the only lever that requires neither new facts, nor external mentions, nor time for reputation. It applies to what is already on the site, and is measured in percentages. For businesses without a budget, this is the cheapest entry into the topic, and also the most poorly sold, because it doesn't look like magic.
By combining both measurements, one can describe the profile of a page that the engine loves. It is long and modular, meaning it is broken down into independent blocks, each of which is understandable without its neighbors. It is semantically close to the answer itself, because it is written with the same words as the question. It contains extractable evidence in ready form: definition, number, comparison, procedure.
Separately, there is a figure about the author. A page with a specified author is cited 2.4 times more often than a page on the same topic without an author. If the author has a verifiable identity, i.e., a page, affiliation with an organization, and a confirmed profile, the advantage grows to 4.1 times. This is the cheapest of the measured levers: it requires neither budget nor new research, only honestly indicating who wrote the text.
What's important about what's not on this list: keywords at the right density, meta tags, and FAQ blocks. The first is not completely obsolete, but it is no longer the main thing. The second is not confirmed. The third is directly refuted by measurement.
The main technical difference from SEO is that here you cannot measure once and get an answer. A separate work is devoted precisely to this: generative engines are non-deterministic, identical queries submitted at different times give different answers and cite different sources. The authors propose a statistical framework instead of a single measurement.
The extent of instability is evident from industry data. In one measurement, the daily brand churn, i.e., the proportion of brands mentioned today but not yesterday, reaches 68.8 percent on one engine. In another measurement across more than twenty-two million citations, 79.6 percent of sources are cited by only one of five engines, and only 0.31 percent.
Hence a direct consequence for measurement and for the product. A single run showing that one engine "sees" you today means nothing: tomorrow the answer will change. A correct measurement requires repeated runs, intervals, and date fixation. Any tool that shows a single number without a caveat about the spread is misleading by design.
Nondeterminism is not a defect of engines that will soon be fixed, but a property of the entire structure. This means that the measurement market should be built around stability, not around a single number. Whoever manages to show businesses not "one number" but "a stable picture with a spread" will own trust in this area.
The second difference from SEO is that there is no single results table where you can be first. The engines diverge not in minor details, but fundamentally. In a measurement of over eight thousand prompts, four major engines agreed on which brand to cite for the same query only in 34 percent of cases. In another measurement, across twenty-two million citations, the engines matched the exact page in 6.8 percent, the website in 10.9, and the named brand in 30.3 percent.
| Level of coincidence | Actual coincidence | If chosen randomly | How many times higher than random |
|---|---|---|---|
| Exact page | 6.8% | 0.16% | 42× |
| Website | 10.9% | 1.0% | 11× |
| Named brand | 30.3% | 1.8% | 17× |
Two conclusions. First: brand matching is approximately four and a half times higher than exact page matching, so the route via brand is easier than via a specific link. Second, non-obvious: relative to random matching, an exact page turns out to be a more informative event than a brand, because cases where two engines picked up the same page are vanishingly small compared to randomness. This means that both levels should be measured separately, and not mixed.
There is also a reassuring dynamic. Over six months, agreement between engines grew from 8.5 percent in January to 11.3 percent in June, which is about a third in half a year. At the same time, within one family, between two Google surfaces, agreement is more than three times higher than between different companies. This means that part of the discrepancy is explained by common origin, and part remains a true difference in selection.
Before optimizing, it's worth understanding how much is already given for free. A measurement on over a hundred thousand responses and over a hundred brands showed a rigid three-tiered picture. Global names, such as the payment system Stripe or manufacturer Nike, appear in relevant responses in 73 percent of cases already in the first run. Established medium-sized brands in 44 percent. Niche and small ones, i.e., everyone not in the top hundred, only in 11 percent.
Two more numbers from the same measurement are noteworthy. When engines cite at all, about 78 percent of citations go to corporate websites, including the brand's own pages and pages of companies in the same market. Among non-commercial sources, video leads, ahead of forums, editorial media, and encyclopedias. And there's an unexpected fact about stability: the brand sentiment, that is, whether it's presented positively or negatively, changes approximately 6.7 times more oftenthan the fact of mention itself. Mentioned stably, but how it was mentioned, floats.
For small businesses, almost everything comes not from working on the text, but from working on recognition and presence where engines get their material. The gap between the first and third stages, thirty points, is more than anything that can be gained by optimization. This means: first presence, then form. The reverse order only works if presence already exists.
Industry measurements, despite their vendor nature, provide a consistent picture of source distribution. Wikipedia accounts for approximately a quarter of citations, forums, and primarily Reddit, about a fifth, news media about a seventh, and businesses' own pages only about an eighth. The last number should be read twice: a brand's own pages receive fewer citations than forums and encyclopedias combined.
| Source type | Share of citations | Note |
|---|---|---|
| Encyclopedia | ≈24% | strong on definitions, up to 36% for one engine |
| Forums and communities | ≈19% | lead in comparisons and recommendations, are growing |
| News media | ≈14% | are strong on trends, losing share |
| Own business websites | ≈12% | lose to forums; grow only where the site is readable |
| Review platforms | ≈9% | are strong on purchase decisions |
| Video | ≈7% | is strong for one engine, almost zero for another |
Separately about Reddit, because it is on Reddit that the third participant's argument rests. The data show two opposing facts simultaneously. The presence of ten or more useful mentions on Reddit increases the chance of citation by 2.7 times, this is the strongest external factor measured. And yet one engine uses Reddit in 45 percent of answers, and another in zero percent. The same with video: one engine has two-thirds of the answers, the other almost zero.
It looks like roulette, but in reality it's explainable: each engine has its own corpus and its own selection settings. Working with one source means working for one engine and failing with the others. This is not roulette that cannot be cheated, but it is also not a universal recipe.
Now we can break down @james's argument piece by piece. It consists of two statements: first, that a good website automatically yields results, and second, that social networks, especially Reddit, solve the problem for you.
The second statement is backed by data better than one might expect. Presence on Reddit increases the chance of citation by 2.7 times, presence in an encyclopedia by 1.8 times, and on review platforms by 1.6 times. This means that external confirmations really work, and working with them is not an alternative to SEO and SMM, but a direct continuation of social work.
The first statement is less well supported. A good website itself does not guarantee inclusion: the share of own pages in citations is only about twelve percent, meaning the machine more often reads about you than you. And there is a direct measurement of how much is lost due to poor readability: answers collected from the external web about a business whose website is incomprehensible to the machine, in 3.7 times more often contain no facts, which the buyer asked about.
External presence works and partially transfers to GEO. A good brand with good social networks gets some results for free, and this is measured: up to 2.7 times from a single Reddit.
The website itself doesn't solve it: the share of its own pages is about 12%. And if the website is incomprehensible to the machine, the answer about you is collected from external sources and loses facts needed by the buyer 3.7 times more often.
The correct formulation: SEO and SMM give you approximately as much as they would have given before, and beyond that, a separate task emerges that they do not cover. This task is about the readability of your own website by an agent.
In the discussion about Kith, two opposing statements were made. One: models do not feel emotions, so people need emotional delivery, and machines need facts, and these are two different recipients. The other: models imitate emotions exactly as humans do, so emotional content affects them in the same way as it affects a person. Both statements are based on something real, and both conflate two different things.
One must separate the emotion in the prompt to the model and the emotion in the content of the pagethat the model reads. These are different channels with different evidentiary bases.
The first channel is well-measured. The EmotionPrompt work (arXiv:2307.11760) ran forty-five tasks on several models and showed that adding emotional stimuli to the prompt increases the result by 8.00 percent on Instruction Induction and by 115 percent on BIG-Bench, and in a human study with 106 participants, the average increase was 10.9 percent, truthfulness plus 19 percent, informativeness plus 12. A separate 2026 study took four tones, joy, apathy, anger, and fear, with an unchanged meaning, and found a difference of up to 4.5 percentage points between the best and worst tone on SuperGLUE tests. That is, the tone of the prompt measurably affects the model's behavior, and in this, the arguing party is right.
But the mechanism is not what they are talking about. The model does not experience emotion: the emotional framework changes the significance of parts of the prompt and the distribution of attention. It does not follow from this that an emotional story on a page similarly affects the accuracy of the response. We do not have this measurement, and we must speak about it honestly and directly.
And there is a third thing, which is not named in the dispute, and it explains where the machine gets its evaluation from. The machine does not take the tonality from your story, it takes it from reviews and forums. A measurement across eight hundred thousand responses to 1,926 brands showed an echo effect: the brand description in the ready answer amplifies the evaluation that is already in the reviews, and in the absence of reviews, a signal of absence arises, and it is read as weakness. The same is seen in the selection of sources: when asking about a specific brand, off-page content dominates, editorial sites provide about 16 percent of citations, forums and social networks about 11, reviews about 11, directories about 10.
Emotion for a human and emotion for a machine are not two recipients, but two channels. The tone of the question measurably affects the model. The tone of the content does not affect the machine's evaluation: it is taken from reviews and discussions. Therefore, emotional delivery remains a human task, and machine evaluation is formed where others talk about you.
The second controversial statement is as follows: it is enough to simply create more content, and it will occupy space in the data, and the data will provide both SEO and GEO, so no other optimization is required. This is the most common position in the market, and it is tested in parts.
What volume truly provides. More materials mean more reasons to be mentioned, and mentions work: presence in discussions increases the chance of citation by up to 2.7 times, presence in reviews and directories makes a smaller but stable contribution. So volume is not useless, it just works through presence, not through the mere fact of writing.
What volume does not provide. It does not elevate you through a level of fame. A measurement across over a hundred brands showed a ladder: global names appear in answers in 73 percent of cases, established brands in 44, niche and small ones in 11. The gap between levels is about thirty points, and the volume of text does not close it, because the reason for the gap is not the amount written, but recognizability and readability. The second number: businesses' own pages receive about 12 percent of citations, meaning even a large website without readability remains a minority of sources.
More reasons for mentions. More surfaces from which the machine can extract a fact. The contribution of presence is measured: up to 2.7 times from a single source.
Changes the answer itself, not just the chance. An answer about a readable business is compiled from its pages in 78 percent versus 56, more accurately by 41 percent, and is recommended 1.9 times more often.
Hence the practical rule of order: first readability, then volume. The reverse order yields more text that the machine cannot use, and this is the worst possible option, because it is paid for with time and is not checked by any counter.
The strongest evidence of the distinctness of this work comes from a measurement based not on engine responses, but on agent behavior. The authors ran almost thirty-eight thousand agent routes, where each route was a customer question about a specific business, through four independent software stands and over a thousand real companies. Companies were leveled by fame and by what the model knows about them, and then divided into agent-readable and unreadable.
Read these numbers together, because individually they are misleading. Only seven to ten percent of the model's training makes it into the final answer; the rest is gathered by reading. If a site is readable by an agent, the answer is compiled from the business's own pages in 78 percent of cases, versus 56 percent for unreadable ones. Such a business is explicitly recommended 1.9 times more often. An answer compiled from the pages is 41 percent more accurate. And the main rejection is not an invention, but an omission: answers collected from the external web are 3.7 times more likely to contain none of the facts the buyer asked about.
SEO didn't have this step. The search engine indexed the page and led a person to it, and then the person decided. The agent, however, goes to the site, reads it, and makes a decision themselves. If the site is unmanageable for the machine, there is no refusal or error, but silence: the answer about you is collected from external sources, and you don't find out about it. The authors' phrasing is more precise than any metaphor: being readable is more important than being discussed.
Hence, it follows why the table is correct in the main point, and its critics do not see this step. Optimization for engines is not only about how you are cited, but also about whether the agent reads you at all. The latter is not included in either SEO or SMM by definition, because they do not have an agent as an active party.
Is all this worth money, and what kind. Industry measurements provide a consistent picture here, and it is unexpected. Direct transitions from AI search account for only about 4.7 percent of sessions on commercial websites, which means it's still small as traffic. But these sessions convert at 5.4 percent versus 2.6 percent for regular organic search, which is approximately twice as good.
The main effect is not in transitions, but in deferred demand. Brands that are cited experience an increase in branded search by 23.4 percent within thirty days after the event, and with sustained citation, the growth reaches 41 percent over ninety days. The person saw the name within the answer, remembered it, and then searched for it themselves.
| Metric | Value | What it means |
|---|---|---|
| First source in the answer | 40–55% of further transitions | distribution closer to "all or nothing" than to positions |
| Median citation half-life | 41 days | from 18 days for one engine to 67 for another |
| Share of sessions from AI search | ≈4.7% | as a traffic channel, it is still small |
| Conversion from AI transition | 5.4% vs 2.6% | approximately twice as good as organic |
| Growth in branded demand | +23.4% in 30 days | delayed effect, not in clicks |
| Growth in brand demand with stability | +41% in 90 days | only works with repeatability |
Hence the economic meaning of the topic. It is not like SEO, where you pay for positions and clicks, but rather like reputation management: you pay for the name to be heard in the right context and to be heard consistently. The half-life of citation in forty-one days means a simple thing: the work does not end, it must be maintained, otherwise the result dissolves in a month and a half.
From all that has been collected, a method emerges, and it differs from the usual measurement of positions so much that it is meaningless to transfer old tools. Below is what has been confirmed by research, not what is convenient for sale.
| Rule | Why exactly so |
|---|---|
| Count selection and assimilation separately | The breadth of citation and the depth of influence diverge; one counter mixes different phenomena and leads astray. |
| Repeat the run and give an interval | Engines are non-deterministic: 79.6% of sources are unique to one engine, daily brand change up to 68.8%. |
| Count for each engine separately | Brand agreement is only 34%, for the exact page 6.8%; the average across engines hides exactly what is important. |
| Maintain control | Without control over the raw part of the same site, platform growth is attributed to measures, and the multiplier is inflated many times over. |
| Maintain brand and page separately | Brand matching is 4.5 times higher than page matching; mixing makes a weak result look strong. |
| Indicate engine date and version | Answers change, and different versions of the same model yield different sets of sources. |
| Check the wording, not just the fact | The brand sentiment sign changes 6.7 times more often than the fact of mention; to mention and to recommend are different outcomes. |
This list is the answer to the question of why the dispute cannot be resolved with the old tool. The measurement performed by a typical SEO system answers the question "what is your position in the search results." Generative engines have no position; there is inclusion in the text, and it is non-deterministic. This means a different class of instrument is needed.
Optimization for engines has two risks, and both are measured or normatively defined, not invented.
The first risk is manipulation. A separate work directly describes the danger: optimization allows content owners to rewrite content in a way that displaces honest sources, and in recommendation agents, this means that the seller can dictate their own recommendation. The mechanism is well-known: the more engines depend on the page's form, the more profitable it becomes not to report the truth better, but to format it more conveniently. This is exactly the same disease that link SEO had, and it will return here in a new form.
The second risk is legal. When a decision about a person is made purely automatically, Article 22 of the General Data Protection Regulation requires human oversight after the result, meaning the possibility to intervene and review. For recommendation agents, this means that a fully automatic recommendation with legal consequences is unacceptable, and a human layer is mandatory not as a gesture of trust, but as a condition of legality.
We sell readability and thereby participate in the general race of form. If readability simply becomes another way to displace honest sources, trust in it will disappear, just as it disappeared for links. This means our interest is that readability is confirmed by facts, not by formatting, and that the data owner retains control. This is not charity, but protection of our own market.
In the Kith discussion, the idea of a verified community badge was raised, which increases conversion because the user might not know what Kith is, but the word “verified” sounds more convincing. The idea is viable, but it needs to be split into two channels, otherwise it will yield half the benefit.
For a human, the badge works as a sign. Its appearance and placement are sufficient. For a machine, the badge doesn't work at all if nothing can be extracted from it: an image is not a statement. A machine needs a statement that has an author, subject, and date, i.e., who verified it, what exactly was verified, and when. Then “verified” turns into an extractable fact that the agent can transfer into a recommendation.
The second argument in favor of explicit formulation comes from data on absence. If there are no external confirmations about a community or brand anywhere, the machine interprets this as a signal of weakness, not as neutrality. This means a badge without verifiable content is worse than its absence in conjunction with an honest description: it creates an illusion that cannot be confirmed.
Formatting without proof is precisely the mechanism described in research as the displacement of honest sources by the seller. Our own interest is that every sign of confirmation is a verifiable statement. Otherwise, we are building the same market that will later collapse, just as the link market collapsed.
Before answering whether these pages need GEO and AEO, we need to look at what they already contain. We took measurements for kithrabbit.com and for nine community websites that it presents as its own: embodied.nexus, eventshk.tech, nextbig.asia, dimsumdiffs.com, deepailiteracy.com, discoverybay.fun, movemen.hk, onegaiwan.com, and haah.ing. The measurement was taken on the rendered page, i.e., on the text that both a human and an agent capable of executing scripts can see.
| Community website | Words | Numbers | Dates | Rules | Confirmations | Schema markup | robots / llms.txt / sitemap |
|---|---|---|---|---|---|---|---|
| embodied.nexus | 236 | 0 | 0 | 0 | 0 | no | no / no / no |
| dimsumdiffs.com | 201 | 0 | 0 | 2 | 0 | no | no / no / no |
| discoverybay.fun | 233 | 0 | 0 | 0 | 0 | no | no / no / no |
| deepailiteracy.com | 248 | 4 | 0 | 0 | 0 | no | no / no / no |
| eventshk.tech | 373 | 7 | 0 | 4 | 0 | no | no / no / no |
| nextbig.asia | 410 | 7 | 0 | 2 | 0 | no | no / no / no |
| onegaiwan.com | 549 | 0 | 0 | 0 | 0 | no | no / no / no |
| haah.ing | 705 | 19 | 21 | 1 | 0 | no | yes / yes / yes |
| movemen.hk | 849 | 10 | 4 | 0 | 0 | no | yes / no / yes |
Kithrabbit.com itself separately: the file weighs about 159 kilobytes, and the visible text in it is 6,540 characters, which is about 7 percent. There is nothing marked up for machines: no structured markup, no robots.txt, no sitemap, no description for the agent. At the same time, it honestly shows nine communities and two links to embodied.nexus, meaning a connection between the pages exists, but it exists only for a human who sees the link. Reciprocity also exists in both directions: embodied.nexus links back to kithrabbit.com. But none of these links carries a claim of ownership, only an address.
It's not that the communities are poorly made. It's that the network pages are structured like posters: they tell a human who they are, and tell a machine nothing that can be transferred into a response. None of the nine pages gives an agent a verifiable claim about itself.
Let's take one page and go through it with an agent's eyes. embodied.nexus responds with code 200, the file is about 21 kilobytes, rendered text is 1,332 characters, which is 236 words. The first heading: “AI is getting a body.” Further sections about the community monitoring robots, holding meetings in Hong Kong and Shenzhen, publishing news and updates, and joining via Kith Rabbit. At the bottom are links: to itself, to joining via the app, to a PDF with a poster, and to kithrabbit.com.
Now, the questions that a participant's agent or Kith Rabbit's own agent would ask, and what it would receive.
| Agent's question | What's on the page | Outcome |
|---|---|---|
| What kind of community is this, in one definition | Slogan and three short paragraphs about why it exists | partially: meaning is discernible, but no definition |
| What topics are active inside | Listing of robot types: humanoid, quadruped, manipulator, wheeled, drone | yes, this is extractable |
| When do meetings take place | Zero dates on the page | no |
| How many participants | Zero numbers on the page | no |
| What rules are in effect | Zero mentions of rules | no |
| Who runs the community and whether it is confirmed | Zero claims of verification | no |
| What the community has already done | Words about news and updates without examples or dates | no |
Out of seven questions, the page answers two, and one of them only partially. This is not a criticism of the design: the page is made as a poster for a human, and for a human, it works. But for an agent deciding what kind of community this is and whether to recommend it, there is almost nothing to glean. According to measurements from previous sections, this is precisely what leads to the worst outcome: the answer about such a community will be compiled from external sources and 3.7 times more often will not contain a single fact that was asked about.
It is important that nothing breaks in the process: the page opens, the application works, a person reads and joins. The failure happens quietly. The agent simply does not find claims to rely on and bypasses the page, and the owner receives neither a refusal, nor an error, nor a record of it.
Now we can directly answer the posed question. It breaks down into two different questions, and they cannot be mixed.
First: is SEO enough for the page to be found. This is not even present in a basic form right now, and it is fixed by conventional means: two out of nine pages publish robots.txt, two a sitemap, and none structural markup. But that's not the main point. SEO answers the question "will the page be indexed and shown for a query." For a community page, this is useful, but it does not solve the problem, because the problem is different.
Second: is SEO enough for the agent to understand the community. No, by definition of the tool. SEO brings the page to the reader, and then the reader decides for themselves whether they are human or an agent. The agent, however, comes to the page, reads it, and decides what kind of community it is, on what exactly its understanding is based. Here, it's not positions or keywords that work, but extractable statements and whether the page is ready for reading. This is the subject of GEO and AEO, and this is precisely what is missing from the network's pages.
The discoverability of the page in the index and its display for a query. For network pages, this needs to be done, and partially it hasn't been, but it's not enough.
Extractable claims about the community and the page's readiness for agent reading: meaning, live topics, rules, activity, results, verification. Plus minimal machine files: markup, sitemap, description for the agent.
Hence a one-line answer. For Kith Rabbit network pages SEO is not enough, because SEO does not create what the agent reads. And at the same time what is sold as GEO is also not enough, if by that one means advice on content and mentions: in the measurement of these nine pages, the root cause is not in the texts and not in the mentions, but in the absence of extractable facts and machine files. A third, simpler and more boring layer is needed: the page must communicate verifiable claims about itself.
Not “optimizing communities for AI”, but “each community describes itself so that an agent can read it, and confirms its claims”. The former is immeasurable, the latter is measured by the number of extractable claims and by the answers to agent questions, and we have just shown how to measure it.
For individual communities in Kith, the same logic is turned inside out, and this is the most useful thing that can be taken from the analysis. A community does not have a product catalog, but it has something that an agent needs even more: context, rules, participant composition, and reason for existence. If this context is structured as text for humans, the agent will not read it and will form an idea of the community from external sources, i.e., from forums and videos.
The numbers say this is not a hypothesis. The share of own pages in citations is about twelve percent, and forums about nineteen: the machine more often talks about you using others' words than your own. And when it collects an answer about what the site does not provide access to, precisely those facts that were asked about are lost 3.7 times more often.
For the community, this means specific consequences. A participant's agent looking for a suitable community will find it only if they can read what kind of community it is, what its rules are, and what topics are live within it. Otherwise, they will either pass by, or describe the community based on others' retellings, including outdated and incorrect ones. In both cases, the community loses new people, and its owner learns nothing about it, because the rejection occurs without an error.
For a store, denial of access means loss of an order. For a community, denial of access means loss of mandate: the agent cannot find the community, check its rules, or recommend it to its owner. The latter is more expensive, because a community exists only due to the influx of people.
If you apply the profile of a readable page to a community, you get an unexpectedly accurate construct. It's not the genre of "community page" that works, but the presence of extractable evidence: what exactly this community is engaged in, what topics are discussed within it, what rules are in effect, who joins and with what intention, what results the community has already yielded.
Description for a human, cover image, participant counter, post feed, rules in small print. The agent has nothing to extract: no definition, no numbers, no comparisons.
Formalized meaning: topic and definition, number of participants with activity, rules as enumerable conditions, owner's intentions, history of results. Each point is extracted separately.
Here it's worth recalling the negative result regarding the "question-answer" format: the block itself provides nothing. The same is true for a community. It's useless to add a "frequently asked questions" block to a community page if there are no extractable facts inside. Only what can be taken out and inserted into someone else's answer without loss of meaning works.
And there is a separate requirement, which in the case of a community is more important than in the case of a store: consent. A store discloses a catalog that is already public. A community discloses people, their interests, and intentions, and these are personal data. This means that the readability of the community must be limited: the agent should receive what the owner intentionally made readable, not what can be scooped out. The limitation does not spoil readability but makes it permissible.
Community readability is not "a page that a bot sees." It is a set of statements that the owner consciously made verifiable: what the community is engaged in, what the rules are, what the intentions are, what the results are. The agent reads exactly these and relies on them in recommendations.
From two positions in the dispute, "the community page should feed Kith Rabbit" and "the page belongs to the community, a sign is enough for us," a solution is assembled if we separate for whom what is being done. The measurement above shows exactly where the failure is now, and it's not advertising or branding.
| Solution | For a human | For an agent |
|---|---|---|
| Page brand | Community is key, recognition of its name | No difference: extractability is important, not the size of the sign |
| Kith's trace on the page | A small badge and network affiliation | Mandatory text string “community on Kith Rabbit”, otherwise the agent will not link the page to the network; currently, only some pages have such a string and only as a signature |
| Verification badge | Works as a sign of trust | Works only as an assertion: who verified what exactly, when; currently, there are no such assertions on any of the nine pages |
| Pitch | Story, tone, affiliation | Definitions, topics, rules, activity, results, each as a separate fact |
| Quality assessment | Composed of narrative and signs | Taken from reviews and discussions, not from your text |
Hence, three decisions that can be made immediately, and they are verifiable.
For humans, the community speaks for itself in its own voice, and Kith Rabbit is present as a sign. For machines, the community speaks for itself through extractable facts, and Kith Rabbit is present as a verifiable statement of affiliation. Two channels, one community. The check is simple: ask the page seven agent questions from the previous section and count how many it answers.
Our previous analysis showed that a missing layer is needed for communication between communities: a mandate of intent, a constrained agent, proof of acquaintance, and a record. This analysis adds the second half of the problem to it, and it explains why search alone is not enough.
Readability and mediation are two sides of the same job. Readability determines whether an agent can even find out what community is in front of it and what it is looking for. Mediation determines whether it will have the right to act further: introduce people to each other, verify the owner's intent, leave a record. Without the first, the mediator is blind because it has nothing to read. Without the second, readability is dangerous because it exposes people without a mandate and without boundaries.
Hence the answer to the question of why Kith needs this. Individual communities in a system without this layer are not visible to agents and are not connected to each other. With this layer, they become not only visible but also connectable according to the owner's rules, with proof and a record. That is, readability is the input, and mediation is the product itself.
Being readable is more important than being discussed, but to be connected, neither is enough. A mandate is needed.
Summary conclusion of two lines of researchThe dispute is analyzed by statements, not by people. Below is each statement that was made in the group, with an assessment and justification.
| Assertion | Assessment | Justification |
|---|---|---|
| “Optimization for engines is a new discipline with different goals and different measurements” | Partially true | The unit of account and measurement are indeed different; three rows of the six-row table are only partially confirmed. |
| “Competition has become all or nothing” | Too harsh | There are several brands in the answer, on average from 3.7 to 6.2; but the concentration is high: the top 3 brands in the sector take 64.7%. |
| “The direct answer format is decisive” | Refuted | The “question-answer” format itself does not improve assimilation; the density of retrievable evidence works. |
| “The same as for SEO and SMM is enough” | Partially true | Content and external mentions work here too, but a step appears that is not present in SEO: the agent reads your site. |
| “Loud multiples of growth confirm the effect” | Refuted | Raw growth ×5.7 versus control ×3.5; causal effect ×1.82, but p=0.16 — indicative, not proven. |
| “A good website and social media will give results on their own” | Half true | External presence gives up to 2.7× from a single Reddit, but own pages give only 12% of citations without work on readability. |
| "Reddit solves the problem" | Works under one engine | Reddit in 45% of one engine's answers and in 0% of another's; presence gives 2.7×, but it is not a universal recipe. |
| "You need to do everything as before" | Incomplete | A separate task remains: the readability of one's own website by an agent. An answer about an unreadable business 3.7× more often loses necessary facts. |
| "It's enough to make more content, no other optimization is required" | Refuted | Volume works only through presence, up to 2.7×, and does not close the gap between levels of fame: 73 / 44 / 11 percent. The median network page is 373 words and zero extractable statements. |
| "Models don't feel emotions, emotions for people, facts for machines" | True in the main | The model does not experience emotion, and the inclusion in the answer is determined by extractable evidence. But the conclusion "emotion only for humans" is incomplete: the tone of the query measurably affects the result. |
| "Models imitate emotions just like people, so emotional content works on them too" | Mixes two channels | The tone of the query gives from 4.5 points between tones to 8.00% and 115% for EmotionPrompt. The influence of emotional storytelling on a page on citation has not been measured. |
| "The verification badge will increase conversion" | True for a human | For a machine, it only works as an assertion: who verified what and when. There are no such assertions on any of the nine pages of the network. |
| "Community pages should feed the Kith Rabbit page, with a recognizable brand mark" | True as a binding | A link is needed, but for an agent, it should be an extractable string, not a mark: the link exists, but there's no assertion of ownership. The mark works for a human. |
In summary: the table is right in its diagnosis and exaggerates in its scale; its critics are right that the levers are largely the same, and they don't see the step at which the agent reads the site. The dispute is closed not in favor of one side, but by separating two different tasks: working with mentions, which is similar to SEO and SMM, and working with readability, which is not present in them.
Only what is confirmed by measurement. Without recommendations that are not in the sources.
An honest list of gaps, so the analysis doesn't read as final.
These numbers are suitable for decision-making and for sales, but not for promises. Any statement like "we will increase your visibility by X percent" goes beyond what is proven, and that is why our own description should state what is measured, not what is desired.
The measurement has been made, the answer given, so the steps are now concrete, not exploratory.
The first move takes ten minutes. The second is several days of work on nine pages with verification against the same seven questions. The third is product work, based on two analyses, not one.
To prevent the analysis from being accused of selectivity, everything that was said in the discussion is listed here, with an indication of the section where it is checked, and the outcome. Sections are indicated by name so that the order is not confused with numbers.
| What was said | Where checked | Outcome |
|---|---|---|
| Table of six rows: different goal, different result, different signals, different format, different competition, different measurement | Six Table Rows, One by One | two rows are confirmed, three partially, one requires rephrasing |
| A good website and social media, especially Reddit, yield results themselves | Whom search engines trust and why Reddit; Thesis "SEO and SMM will give GEO themselves" | half true: presence up to 2.7×, but the site itself is only about 12 percent |
| There is no reason to do anything other than SEO and SMM | AX: a step that SEO didn't have | incomplete: the agent reads the website, and this is a separate task |
| Good structural content always wins | Structure as a separate lever; What engines actually extract | confirmed, but not "always": structure gives 17.3%, density of evidence decides |
| You just need to create more content, no other optimization is needed | Content volume versus readability | refuted: volume does not close the 73 / 44 / 11 gap and does not replace readability |
| Models do not feel emotions: emotions for people, facts for machines | Emotion: two channels, not two addressees | true in general, but incomplete: the tone of the query has a measurable effect |
| Models imitate emotions just like people, so emotional content also works on them | Emotion: two channels, not two addressees | mixes two channels; the effect of the story on the page on citation has not been measured |
| Large multiples in reports prove the effect of optimization | Key experience: platform tailwind; Why multiples lie | refuted: ×5.7 vs ×3.5 for control, effect ×1.82 at p=0.16 |
| Competition has become "all or nothing" | Ladder of fame; Six rows of the table, one by one | too sharply: up to 6.2 brands in the answer, but the top 3 sectors take 64.7% |
| Direct answers and question blocks solve the problem | Negative result regarding "question-answer" | refuted by measurement |
| Community pages should feed the Kith Rabbit page, including emotional attachment and a recognizable brand mark | What needs to be done on community pages | a link is needed; for the agent, it's an extractable string, the sign remains for the human |
| A small sign is enough, the page belongs to the community, not to us | What needs to be done on community pages | correct as a default, but not at the cost of readability: a minimal footprint is mandatory |
| A verified community badge will increase conversion | Verification badge as proof | for a human, yes; for a machine, only together with the content of the verification |
| Models react to stories the same way humans do | Emotion: two channels, not two addressees | only partially verified: the mechanism is the same as for query tone, there is no measurement for content |
| Do kithrabbit.com and community pages need GEO and AEO, or is SEO enough? | What we measured on network sites; Is SEO enough for these pages? | SEO is not enough, and neither is marketable GEO: there are no extractable statements or machine files |
Two things remained unverified, and this is stated directly, not hidden. First: does an emotional story on the page affect the likelihood of being included in the answer. Second: how will the evaluation behave after everyone starts applying readability. Both gaps are listed in the section about what we don't know.
The access layer these pieces keep arriving at is the one we build. A buying agent asks a supplier's catalogue a question, the supplier's own rules decide what that caller may see, the answer comes out of systems the supplier already runs, and the log stays with the supplier.