Ranking decides most of the outcome because ranking is what determines whether your page is in the pile the model reads at all: every major answer engine retrieves before it writes, and context position is one of only two levers the 2026 critical survey grades at high confidence.2 Relevance and ranking carries the controlled measurement of how large that advantage is and what its limits are; the consequence for this page is only that a team reallocating effort from ranking to rewriting is trading a strong lever for a weak one, which is exactly the split the GEO versus SEO stage argues against.
GEO versus SEO · Topic
What carries over from SEO to GEO, and what does not?
Three fundamentals carry straight over and still decide most of the outcome. The 2026 academic audits show why the honest phrasing is “necessary and increasingly insufficient” rather than “rank first and citations follow”, and why the overlap figure has to be quoted as a dated range.
On this page
The short answer
Three things carry straight over from SEO and together account for most of the outcome: relevance and ranking, crawlability and indexation, and topical coverage. What stops carrying over is the measurement layer, not the content or the technical work. The 2026 critical survey grades query relevance and context position as major determinants at high confidence, while academic audits put the share of domains a Google AI Overview draws on that sit outside the organic results at roughly 30% to 53%, depending on sample size and collection date.239
Key takeaways- Ranking is necessary and increasingly insufficient: it is the strongest lever anyone has measured, and it has stopped being a guarantee of citation.
- Quote the organic overlap as a dated range, never a point: 53% of consulted domains outside the top 10 in September 2025, 29.8% outside the whole first page in March and April 2026.
- The denominator moves the number as much as the engine does: per-answer and per-citation readings of the same data differ by a factor of two or more, and both are accurate.
- Markup is not the lever it is sold as. Google states there is no special schema.org structured data you need to add for its AI features.
Continuity
Which three things carry straight over from SEO?
Relevance and ranking, crawlability and indexation, and topical coverage. Each has its own page in this stage, because each has its own evidence base.
01
Relevance and ranking
Being the best match for the question is still the dominant signal. C-SEO Bench, presented at NeurIPS 2025, reports that making a document first in the model’s context “leads to far greater citation ranking gains in the LLM response than any C-SEO method”, and only 3 of its 54 method–domain combinations came out significantly positive.12 Read the full study.
02
Crawlability and indexation
Google states the requirement in one sentence: “a page must be indexed and eligible to be shown in Google Search with a snippet.”5 There is no separate AI index to submit to, and 27.1% of the URLs AI answers already cite cannot be scraped at all.8 Read the prerequisite in full.
03
Topical coverage
Fan-out makes breadth pay more than it did. Google confirms both AI Overviews and AI Mode “may use a query fan-out technique” that issues “multiple related searches across subtopics and data sources”, so a page answering a whole cluster is eligible at several retrieval points.5 Read why breadth pays.
The strongest evidence
Why does ranking still decide most of the outcome?
The honest limit
Where does ranking stop being enough?
Ranking stops being enough somewhere in the long tail, and the size of the gap is a range rather than a number: between roughly 30% and roughly 53% of the domains a Google AI Overview draws on were outside the organic results for the query that triggered them, depending on which academic sample and collection date you take. Kirsten et al. (Findings of ACL 2026) audited 4,706 queries collected in September 2025 in the US and Germany and reported that “on average 53% (27%) of domains that AIO consults are not contained in top-10 (top-100) Organic search results”.3 Xu, Iqbal and Montgomery measured 55,393 trending queries collected between 13 March and 21 April 2026 and found that “29.8% of AIO reference domains do not appear anywhere on the corresponding first page”.9 The later study is about twelve times larger and roughly halves the figure; the earlier one says consults, which is not identical to cites. Quote both dates or neither.
A third audit says it sideways. Grossman et al. (SIGIR 2026) measured URL-level Jaccard similarity of 0.11–0.18 between organic Google results, AI Overviews and Gemini across 11,500 real user queries: three surfaces of one company disagreeing about which URLs answer the same question.4 The mechanical explanation is fan-out rather than a secret ranking factor: the visible query is decomposed into sub-queries the searcher never sees, each running its own retrieval, so a page can be cited while ranking nowhere for the phrase actually typed. That does not contradict the section above, because the pages supplying high-frequency, repeatedly reused citations still overwhelmingly rank, while the one-off long tail increasingly does not.
The defensible sentence is that ranking is necessary and increasingly insufficient. “Rank first and citations follow” is too strong, and “ranking no longer matters” is contradicted by the graded lever above and is the more expensive mistake of the two.
Reading the numbers
Why does the organic-overlap number keep changing?
It keeps changing because it is defined by a denominator and a date, and published figures move by a factor of five when either one changes. They run from roughly a sixth to over nine-tenths, and almost all of that spread is bookkeeping rather than disagreement about the world.
The denominator moves the answer more than the engine does. “What share of AI Overviews contain at least one URL that ranks in the top 20?” is a per-answer question, and because an overview cites several sources it resolves in the nineties. “What share of individual citations come from a page ranking in the top 10?” is a per-citation question over the same data, and it resolves far lower. Same engine, same week, two numbers that differ by a factor of two or more, both accurate. A figure quoted without saying which of those two questions it answers is not a measurement, it is a decoration.
Time moves it again, and the two academic audits above are the cleanest demonstration: nearly the same construct, seven months apart, disagreeing by half, with the larger and later sample landing at the low end.39 Neither is wrong. They are two readings of a system that changed in between, on numerators that differ: one counts what the overview consults, the other what it references.
Vendor time series point the same way with the same caveat. One study of 863,000 keyword SERPs put citations from top-10-ranking pages at about three-quarters in July 2025 and about two-fifths in March 2026, reading the move as a real change in how Google selects sources rather than a methodology change. Another vendor published two irreconcilable figures for the same nine industries five months apart, roughly 55% and roughly 17%, without acknowledging the contradiction. Vendor-published; described rather than linked, per the curriculum's sourcing rule.7
So quote the range and the reason, never the point: in academic samples collected between September 2025 and April 2026, roughly 30% to 53% of the domains a Google AI Overview drew on sat outside the organic results for the triggering query. If that will not fit on the slide, the slide is asking for a number that does not exist.
Markup and files
Do schema markup and llms.txt files carry over?
They keep the value they always had, and nothing more. The AI-specific upgrade of the claim is what first-party documentation refuses.
Structured data still earns ordinary rich results in classic search, and that is worth doing on its own terms. What does not survive contact with the documentation is the upgraded pitch, that schema markup lifts citation rates in AI answers. Google’s own page on AI features says the opposite in plain words: “You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add.”5
The llms.txt question resolves the same way and more sharply. The claim that an llms.txt file helps a page get cited is unsupported by anything first-party, and Google's AI optimization guide lists “LLMS.txt files and other 'special' markup” among the things to ignore, stating that “you don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn't use them.”10 No other major engine documents reading one either. That does not make the file harmful, just a rounding error next to an indexation problem nobody has looked at.
The rule the two cases share is worth more than either verdict: something carries over from SEO when it was already an input to retrieval. Indexation, server-rendered delivery and topical breadth are inputs, so they carry over intact. A file no documented crawler reads is not, and calling it AI-native does not make it one. That is the cheapest test to apply to whatever tactic is sold next: ask which stage of the pipeline it changes, and if the answer is none of them, it changes nothing.
Discontinuity
What does not carry over?
The content and technical work survives the transition almost intact. The measurement layer and the reporting layer do not.
Carries over unchanged
- Indexation, server-rendered delivery, site speed and URL hygiene
- Topic-cluster work, because fan-out rewards breadth more than it used to
- Maintaining and updating pages you already have
- Earning genuine third-party coverage, which most commercial answers lean on
- Schema for ordinary rich-result reasons, claiming nothing more for it
Does not carry over
- Rank tracking as a measurement method: answers are a distribution, not a position
- The click as the primary KPI; Pew measured ~1% clicking a source inside an AI summary6
- Control of your own title and snippet, since you are paraphrased instead
- One blended visibility score; overlap between surfaces is low enough to report per engine
- Publishing volume as a growth lever, since thin pages compete at every fan-out point and win none
The honest limit of this page
“Necessary and increasingly insufficient” is a direction, not a coefficient. Nobody has published a controlled experiment that moves the same page's organic rank and measures the change in its citation rate, which is the study that would settle how much ranking is worth in an answer engine. Every overlap figure here, academic ones included, is observational: it records that cited pages tend to rank, not that ranking caused the citation. And all of them describe Google surfaces most heavily, because Google is the only engine whose organic results can be compared with its AI output at all.
Where a product fits, and where it does not
The method here needs no software: list the twenty questions your buyers actually ask, check that each has one indexed, well-ranking page behind it, then run those twenty as prompts on each engine several times over a week and record which sources came back. Where the cited source is not you, that is your off-site backlog; where nothing of yours ranks, that is your SEO backlog, and it comes first. Bavior automates only the sampling half, a fixed prompt set across five engines on a schedule with cited sources recorded per run, and it does not do the part this page argues matters most: it will not improve a ranking, fix indexation or build topical coverage. The free GEO audit covers the baseline; paid plans are from $99/mo billed monthly, or $79.17/mo billed annually (as of 30 Aug 2026).
Sources, all checked 30 Aug 2026
- Puerto, Gubri, Green, Oh, Yun, “C-SEO Bench: Does Conversational SEO Work?”, NeurIPS 2025 Datasets & Benchmarks Track; §6.3, placing the target document first “leads to far greater citation ranking gains in the LLM response than any C-SEO method”: arxiv.org/abs/2506.11097
- “Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023–2026)”, 15 Jul 2026, arXiv:2607.14035 (preprint); Table 5 grades “query–document relevance and context position are major determinants” high, and §7.3 gives C-SEO Bench as ~1,900 queries, 16,360 documents, three of 54 method–domain combinations significantly positive: arxiv.org/abs/2607.14035
- Kirsten et al., “Characterizing Web Search in The Age of Generative AI”, Findings of ACL 2026; 4,706 queries collected September 2025, US and Germany. “On average 53% (27%) of domains that AIO consults are not contained in top-10 (top-100) Organic search results”: aclanthology.org/2026.findings-acl.526
- Grossman et al., “How Generative AI Disrupts Search”, SIGIR 2026; 11,500 user queries; “Jaccard similarities between 0.11 and 0.18” across organic Google, AI Overviews and Gemini: arxiv.org/abs/2604.27790
- Google Search Central, “AI features and your website”, last updated 10 Dec 2025 (first-party; indexation requirement, query fan-out, schema.org statement): developers.google.com/search/docs/appearance/ai-features
- Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results”, 22 Jul 2025; 900 US adults, 68,879 Google searches, March 2025; a link inside the summary was clicked “in just 1% of all visits”: pewresearch.org
- Overlap time-series: an AI-visibility vendor study of 863,000 keyword SERPs and ~4M cited AI Overview URLs, published 2 Mar 2026, reporting ~76% of citations from top-10-ranking pages in Jul 2025 against ~38% in Mar 2026. A second vendor published ~54.5% (Sep 2025) and ~17% (Feb 2026) for the same nine industries. Vendor-published; described, not linked.
- Allaham & Diakopoulos, 2026; 27.1% of URLs cited in AI answers could not be scraped. Cited via the critical survey, §8.3; preprint.
- Xu, Iqbal & Montgomery, “Measuring Google AI Overviews”, arXiv:2605.14021 (preprint); 55,393 trending queries, 13 Mar to 21 Apr 2026; “29.8% of AIO reference domains do not appear anywhere on the corresponding first page”: arxiv.org/abs/2605.14021
- Google Search Central, “Google’s Guide to Optimizing for Generative AI Features on Google Search”, last updated 10 Jul 2026 (first-party; the LLMS.txt statement): developers.google.com/search/docs/fundamentals/ai-optimization-guide
FAQ
Frequently asked questions.
Do I still need SEO if I am doing GEO?
Yes, and it is the larger half of the work by the only controlled measurement available. C-SEO Bench, presented at NeurIPS 2025, reports that making a document first in the model's context "leads to far greater citation ranking gains in the LLM response than any C-SEO method", and the 2026 critical survey summarises the same benchmark as ~1,900 queries and 16,360 documents in which only three of 54 method-domain combinations came out significantly positive. Relevance, indexation and topical coverage all sit upstream of anything a generative engine does to your text.
Does ranking first guarantee an AI citation?
No, and the 2026 academic audits show the gap clearly, as a dated range rather than one number. Kirsten et al. (Findings of ACL 2026) audited 4,706 queries collected in September 2025 and found 53% of the domains AI Overviews consult absent from the organic top 10, 27% absent from the top 100; Xu, Iqbal and Montgomery, on 55,393 queries collected between 13 March and 21 April 2026, found 29.8% of reference domains absent from the whole first page, about twelve times the sample and roughly half the figure. Grossman et al. (SIGIR 2026) separately measured URL-level Jaccard similarity of 0.11–0.18 between organic Google, AI Overviews and Gemini across 11,500 queries. The mechanism is query fan-out: sub-queries you never see run their own retrievals, so a page can be cited without ranking for the phrase typed.
What percentage of AI citations come from the organic top 10?
There is no single defensible percentage, because the published figures range from roughly a sixth to over nine-tenths depending on the denominator and the month. "What share of AI Overviews contain at least one top-20 URL" is a per-answer question and lands in the nineties; "what share of individual citations come from a top-10 page" is a per-citation question over the same data and lands far lower. Time matters as much: the two academic audits available disagree by half, with 53% of consulted domains outside the organic top 10 on queries collected in September 2025 and 29.8% of reference domains outside the whole first page on a twelve-times-larger sample collected in March and April 2026. Quote the range, the denominator and the collection date, never one number.
Does adding schema markup or an llms.txt file help?
Schema helps for the ordinary reasons it always did, and the AI-specific version of the claim is not supported by first-party documentation. Google's page on AI features states that "you don't need to create new machine readable files, AI text files, or markup to appear in these features" and that "there's also no special schema.org structured data that you need to add". Its AI optimization guide names "LLMS.txt files and other 'special' markup" among the things to ignore, because "Google Search itself doesn't use them". No other major engine documents reading an llms.txt file. Keep structured data for rich results, and spend the saved hour on indexation instead.
Which SEO habits should I actually stop?
Three, and none of them are content or technical habits. Stop treating a single prompt run as a measurement: the 2026 critical survey reports daily source-level Jaccard of roughly 0.34–0.42 across four engines over 45 days, so a spot check is a sample of one. Stop forecasting sessions from citations: Pew tracked 68,879 searches in March 2025 and found a link inside an AI summary was clicked in just 1% of visits to those pages. And stop reporting one blended AI visibility score, because URL overlap between surfaces of the same company runs at Jaccard 0.11–0.18.
Should I publish more pages to cover more fan-out queries?
Depth on one page beats volume across ten, on the evidence available. Query fan-out issues several related searches per question, and one page that genuinely answers a question plus its obvious follow-ups is eligible at several retrieval points, while ten thin pages each compete at one and win none. The 2026 critical survey reports an end-to-end benchmark over 171,003 documents and 2,700 queries in which body-only optimisation that dilutes topical relevance reduced average top-20 presence by about 9%, so splitting a strong page into fragments carries real downside.
Bavior Editorial
The team that researches and maintains Bavior’s writing on Reddit marketing and AI search visibility. Every figure here is attributed to a named source with the date it was checked, and none of our links are affiliate links.
Found a number that looks wrong? Tell us and we will re-check it: support@bavior.com