What actually gets a page into an AI answer: eight tactics, ranked by evidence

How is the evidence ranked here?

By what kind of test produced it, not by how confident the publisher sounded. Four tiers:

Documented behaviour means the engine says in its own documentation what it does. It is the strongest thing available in this field, because it describes a mechanism rather than an association, and it is also the narrowest: documentation says what gates inclusion, not what wins a citation.

Controlled or matched tests mean someone changed one thing and measured against a control. There are three of these in the whole field that we could read, and two of them returned no effect.

Large-sample correlation means someone counted things that co-occur. It can rule tactics out when it finds nothing and cannot establish that a tactic works.

Vendor assertion means a company that sells the remedy published the claim without a dataset behind it. Nothing in that tier made this ranking.

One more thing belongs in the open: of the eleven studies cited below, nine were published by companies that sell SEO software, AI-visibility tracking or agency services. We say so on each one. We did not exclude them, because excluding them would leave almost nothing, and that is itself the most important fact about the state of this evidence.

The ranking

Rank	Tactic	Best evidence	What it supports
1	Let the crawlers in; be indexed and snippet-eligible	Provider documentation	A prerequisite. Blocking the retrieval bot removes you from the answer
2	Freshness, and a visible recent date	One lab experiment, one observational study	A displayed date moves a model’s choice in a testbed. Production effect unmeasured
3	Third-party mentions, especially on YouTube	Correlation on 75,000 brands	Co-occurrence, nothing causal, and it measures mentions rather than links
4	Surfacing the core topic early on the page	One lab experiment	Topic match gates citation. Formatting does not
5	Ranking in the classic top 10	Descriptive, 4M citations	Helps, and is not required. About a third of cited pages rank nowhere
6	Answer-first formatting, chunking, FAQ blocks	One lab experiment	Null. Formatting-only edits had no measured effect
7	Schema markup	Matched quasi-experiment, 1,885 pages	Null, with a small negative on one surface
8	llms.txt	Provider statement plus two log studies	Disconfirmed. Google ignores it; almost nothing fetches it
9	Paying for an ad inside the answer	Two vendor datasets	The ad does not buy the citation

Nine lines for eight tactics, because the structure tactic splits cleanly into the half that worked and the half that did not. Rows 6 to 9 are the useful part of this table: they are the things a reader can stop doing, or at least stop paying for, on the strength of a real measurement.

Which tactic is actually documented?

Crawler access, and for Google, index eligibility. This is the only item on the list where the engine states the mechanism in its own documentation.

OpenAI publishes what each of its user agents does and says the settings are independent of each other: “a webmaster can allow OAI-SearchBot in order to appear in search results while disallowing GPTBot to indicate that crawled content should not be used for training OpenAI’s generative AI foundation models.” On the consequence of blocking the search bot it is explicit: “Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links.” (OpenAI, Overview of OpenAI Crawlers)

Google’s guidance for its own generative features sets the equivalent gate on the index rather than on a dedicated bot: the page has to be indexed and eligible to be shown as a snippet. (Google Search Central, Optimizing your website for generative AI features on Google Search)

Two things follow, and they point in opposite directions. The first is that this is the one tactic where failure is certain rather than probable: block the retrieval bot and you are out of that engine’s answers, by its own account. The second is that it is a floor and not a lever. Being fetchable is what every page that gets cited has in common, and so is every page that does not.

What it costs to block is a separate question with thinner evidence. A preprint by Zhao and Berman finds “a 7% post-blocking decline in weekly visits measured by SimilarWeb or Semrush within the 6 weeks after blocking”, with a similar estimate from Comscore’s human browsing panel. (Strategic Response of News Publishers to Generative AI, arXiv:2512.24968) Read the base before you use the number: these are news publishers, 24 of which blocked, and 19 of those “made no changes beyond adding GenAI-specific user-agent directives”. The identification window runs from November 2022 to May 2024, which is before Google launched AI Overviews, and the directives studied are training-bot directives rather than search-bot ones. It is the best available estimate of a cost and it is not an estimate of today’s cost. The headline figure also moved between versions of the paper, from 10% in version 3 to 7% in version 4, which is worth knowing if you have seen the earlier number quoted.

On crawl-compensation programmes, where a publisher is paid per crawl or per contribution, we found no evidence either way about whether joining one changes what gets cited. That is a gap, not a finding.

Does a page have to rank to get cited?

No, and ranking still helps. Ahrefs matched 4 million AI Overview citations against the ordinary results for the same query across 863,000 keyword result pages: “37.9% of URLs cited in AI Overviews also appeared within the first 10 blocks. The rest were almost evenly split between positions 11-100 (31.2%) and beyond the top 100 blocks (31.0%).” (Ahrefs, Update: 38% of AI Overview Citations Pull From The Top 10, 2 March 2026)

So a third of the pages Google cites in its own AI answers do not rank in the top 100 for the query the answer responds to. Ranking is a strong tailwind and not a gate. The practical consequence is a measurement one: if you track only your top-10 positions, you cannot see roughly two thirds of the places your citations come from or fail to come from.

Three cautions. This is Google AI Overviews only, and nothing here transfers to ChatGPT or Perplexity. It is descriptive, so it tells you where cited pages happen to rank and not what would happen to a page that moved. And Ahrefs sells an AI-visibility tracker, so the finding that rank tracking is insufficient is a finding its product benefits from.

One number from this study that circulates as a trend should not be: Ahrefs’ own comparison of “~76% in July 2025 vs. ~38% today” is not a like-for-like series. Ahrefs says the parsing changed and the earlier study used only the top three citations per response. Two numbers from different methods are not a slope.

Does freshness work, and does re-dating a page work?

Freshness is the one tactic with a controlled result pointing the same way as the field data, and the controlled result comes from a laboratory rather than a live engine.

The lab study is peer-reviewed, which is rare here. Across “252,000 trials” on six models, with each trial injecting two candidate sources differing in exactly one factor, four factors were decisive in every model tested: “Topic Mismatch, Price Not Mentioned, Recent vs Old Timestamp, and Lower List Position.” (What Gets Cited: Competitive GEO in AI Answer Engines, SIGIR ‘26, arXiv:2605.25517) A visible date changes which of two otherwise identical sources a model cites.

Read the method before you price that in. The date contrast the paper tested is “Content dated 2026 versus 2019”, which is about as wide as a date gap gets. The testbed puts two documents in the model’s context, where the authors note that “Production RAG often retrieves five to ten or more pages, so real citation pools are larger than our testbed”, and they add that production systems “may still favor trusted domains or strong brands when they pick sources”. The paper’s odds ratios are enormous, several above 100. Those are testbed numbers, and treating them as the size of a real-world effect would be a mistake.

The field data is Seer Interactive’s, and it is the single most quoted statistic in this corner of the industry. Of cited pages where both dates could be read, 72% had been updated within the last year, while only 42% had been first published within the last year. (Seer Interactive, Study: Content Recency’s Impact on AI Visibility in 2026, 24 July 2026) The base is “4,124 cited pages where both publish and update dates were readable (ChatGPT, Gemini & Perplexity, Mar-Jun 2026)”, drawn from four client brands.

The gap between those two shares is the interesting part, and Seer puts it plainly: “The freshness LLMs reward is being manufactured by updates, not by new publishing.” What the study cannot tell you, and does not claim to, is whether updating a page causes it to be cited. It observed which pages were cited and when they had last changed. Nobody has run the obvious experiment of re-dating a set of pages on a live engine without changing their substance, which is also the experiment most people asking this question actually want. We go through this in more detail in the freshness illusion.

Do third-party mentions help?

The correlations are the strongest in the field and they are correlations. Ahrefs measured 75,000 brands against brand mentions in AI answers and reports that “YouTube mentions show the strongest correlation with AI visibility (~0.737), outperforming every other factor across ChatGPT, AI Mode, and AI Overviews”, with branded web mentions at 0.66 to 0.71 and almost nothing for the number of pages on a site, at “~0.194”. (Ahrefs, Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews, 12 December 2025)

Ahrefs prints the caveat itself: “The usual disclaimer applies: correlation isn’t causation.” Two more belong next to it. The study measures brand mentions inside answers, which is a different quantity from a linked citation of your page, and the two are easy to conflate when planning work. And the direction of the arrow is genuinely unsettled: brands that get talked about on YouTube are brands that are already well known, and being well known is exactly what would independently raise the odds of appearing in an answer.

There is one intervention study in this area, and its subject was invented. An Ahrefs researcher created a fictional company, planted three false stories about it, and recorded how often models repeated them. It demonstrates that planted third-party sources can enter answers. It involved one fake brand and 2025-vintage models, and it measured the content of answers rather than citation counts, so it supports a claim about mechanism and not a claim about how much a real brand should spend on digital PR.

Does answer-first structure and chunking help?

This is the clearest null result in the set, and it comes from the same peer-reviewed experiment as the freshness finding, which makes the contrast between the two internally consistent rather than a comparison across studies.

The paper tested two formatting factors, defined as “Dense paragraph format versus organized sections” and “Related information dispersed throughout text”. Its result: “Formatting choices (Content Structure, Scattered Information) had no impact, suggesting LLMs parse content regardless of visual organization.” Its abstract states it as “formatting-only edits have little impact”, and its recommendation is that “Formatting changes showed minimal return and can be deprioritized.”

Google’s guidance says the same thing from the other direction, listing among the things you do not need to do that “you can ignore tactics like ‘chunking’ content”.

Now the half that did work. In the same experiment, Topic Mismatch is one of the four decisive gatekeepers, and the paper’s practical advice is to “surface core topic terms early”. That is not a formatting instruction. It is an instruction about what the opening of the page is about.

A position study is often quoted alongside this and measures something different again: Surfer reports that across “over 100,000 AI citation placements across more than 10,000 AI Overviews responses”, “38% of all AI citations are pulled from the first 100 words of a page.” (Surfer, Why Almost 40% of AI Citations Come from Your First 100 Words, updated 17 September 2026) It is vendor-published, Google AI Overviews only, and it records where in a page the cited passage sat. It does not show that moving a passage to the top would get it cited.

Put the three together and the honest instruction is narrow: say what the page is about, early, because topic match is doing the work. The heading shapes, the question formats and the FAQ blocks are not where the measured effect is.

Does schema markup help?

The best test available says no, with a small negative on one surface.

Ahrefs “tracked 1,885 web pages that added JSON-LD schema between August 2025 and March 2026, matched them against 4,000 control pages, and measured citation changes across Google AI Overviews, AI Mode, and ChatGPT.” The results: “Google AIO -4.6% Small but statistically significant decline relative to matched controls”, “Google AI Mode +2.4% Statistically indistinguishable from zero”, “ChatGPT +2.2% Statistically indistinguishable from zero”. On the first of those it reports that “the odds of seeing a gap this large by pure chance are about 1 in 2,500”, amounting to “an average loss of around 12 daily citations per page, in a sample where most pages were getting hundreds”. (Ahrefs, We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved., 11 May 2026)

This is the strongest study design in the whole ranking: a matched before-and-after on live pages rather than a count of co-occurrences. It carries one structural limit that matters for how you read the negative. The sample is pages that already had substantial AI Overview citation volume, so it answers the question “does adding schema increase citations to an already-cited page” and says nothing about whether markup helps a page that is cited by nobody.

On the provider side, Google’s guidance says no special markup is needed for its generative features. We looked for a counterpart statement from another provider recommending schema for LLM retrieval and did not find one in documentation. The claim circulates, attributed to a conference remark, and we could not source it to anything a provider publishes and maintains. That is where it stays until someone can.

None of this is an argument against structured data. It earns classic rich results, and that case does not depend on AI retrieval at all. It is an argument against funding a schema programme out of a budget justified by AI citations.

What has been actively disconfirmed?

Two tactics, and they are the two with the loudest marketing behind them.

llms.txt is the first, and it is the only item in this ranking that a provider has explicitly ruled out in writing. Google: “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.” It goes further and closes the hedge most proponents fall back on: “Doing so will neither harm nor help your site’s visibility or rankings in Google Search, as Google Search ignores them.”

Two log studies from different directions agree. Ahrefs checked every one of “137,210 domains in Ahrefs Web Analytics that received traffic in May 2026”, found that 28% publish an llms.txt, and then found that “97% of those files received zero traffic in May 2026. Nothing fetched them at all.” (Ahrefs, We Analyzed 137K Sites: 97% of llms.txt Files Never Get Read, 15 June 2026) Ahrefs calls its own 28% an upper bound, because its customers skew technical. From the citation side rather than the request side, Allmo found a single llms.txt page among 94,614 cited URLs. (Allmo, llms.txt, 23 January 2026) We covered that count when it appeared, in llms.txt does not work, and the measurement has not improved for the file since.

Paying for an ad inside an AI product is the second. SE Ranking analysed “50,006 commercial prompts across 20 niches in the US” and found that “In 96.37% of placements, the advertiser is not cited among the answer’s sources”, with the exact advertised URL appearing “in only 0.09%, or 12 cases”. (SE Ranking, ChatGPT Shows Ads for 1 in 4 Commercial Prompts, 10 August 2026) A second vendor, using a different sample of 36,153 US answers, puts the overlap at a similar level: “Only 2.8% of ads pointed at a domain the same answer also used as a source.” (LLM Pulse, ChatGPT ads study)

On Google’s surface the overlap is higher and the conclusion holds. SE Ranking reports advertiser domains among cited sources in 11.53% of cases and the exact URL in 1.95%, and describes a matched comparison: “We compared advertisers against non-advertising domains of the same strength for the same query, matched on Domain Trust, backlinks, referring domains, and organic standing. Even then, advertisers were cited no more often.” (SE Ranking, Google AI Mode ads) That page prints no sample sizes or test statistics for the matched comparison, so take it as a vendor’s matched observation rather than an experiment.

One figure we are deliberately not giving you is how often ChatGPT shows an ad at all. SE Ranking says 25.94% of commercial prompts; LLM Pulse says 71.9% of US answers. Those two cannot be reconciled from the published pages, so neither belongs in a sentence on its own. Our earlier piece on ChatGPT ads reaches the same place on the citation question.

What remains untested, and why?

Four things, and the reason is the same in every case: the test requires controlling a live engine’s inputs over time, and the people who could run it either sell the remedy or sell the engine.

Nobody has re-dated a set of pages on a production engine while holding their substance constant. That is the experiment that would turn the freshness correlation into a causal claim, and it is the one most commonly implied by advice to “refresh your content”. An agency could run it on client pages and would then be publishing a result that might reduce demand for content refreshes.

Nobody has added schema to pages with zero AI citations and watched a matched control group. The one matched study starts from already-cited pages, which excludes the population most likely to benefit.

Nobody has tested whether a crawl-compensation programme changes citation selection, as opposed to changing payment. The programmes are young and the data sits with the platforms running them.

And nobody has published a cross-engine replication of any of this. Every study above covers one or two surfaces. Google AI Overviews is heavily measured because its citations are scrapable at volume; ChatGPT and Perplexity are measured through small prompt panels; Claude and Copilot barely appear. A tactic that works on one retrieval pipeline has no claim on the others, and the field currently writes as if it does.

What should a reader actually do?

Make sure the retrieval bots can fetch the page and that it is indexed and snippet-eligible. That is the only step on this list with a documented mechanism and a certain failure mode.

Say what the page is about in its opening, using the words the topic is actually called. Topic match is the one content factor that survived a controlled test, and it is not the same instruction as formatting for machines.

Keep pages genuinely current, and do not expect a date change to do work on its own. The evidence supports currency and does not support re-dating.

Stop funding llms.txt, and stop buying ads to win citations. These are the two places where a real measurement says the money does nothing, and the measurements come from four separate datasets.

Measure your own citations rather than borrowing anyone’s averages, including the ones above. Our method for that needs no paid tool and makes its own confidence interval visible, which most vendor dashboards do not.

Methodology limits of this ranking

Every source linked above was fetched during the writing of this post and returned HTTP 200, and every quoted string was matched against the served text. Four limits are worth stating plainly.

Nine of the eleven studies cited are published by companies selling SEO software, AI-visibility tracking or agency services. We have labelled each one, and the labelling does not neutralise the problem. There is no independent, funded, replicated body of work on this question, and a ranking built from this literature inherits that.

The tiers rank study designs, not effect sizes. A documented provider statement sits above a correlation on 75,000 brands even though the correlation covers more of the web, because a mechanism you can act on beats an association you cannot. A reader who cares mainly about magnitude should re-sort the table and will reach different conclusions.

Several claims in circulation are absent here because we could not source them. A figure we previously reported for how many cited pages rank outside the top 100 is one of them: the report it was attributed to is no longer retrievable, and the number as it circulates describes something different from what we said it did. The figure in the ranking section above, from Ahrefs, is the sourced replacement, and it is a different measurement rather than a correction of the same one. We will correct that earlier post separately.

Finally, this is a snapshot of a moving target. Ahrefs attributes a large change in its own ranking numbers to AI Overviews moving to a new model in January 2026. A ranking of tactics against engines that are re-trained and re-plumbed every few months has a short shelf life, and the right way to use it is as a map of what has and has not been tested, rather than as a settled list.

Sources

Ahrefs, Update: 38% of AI Overview Citations Pull From The Top 10, 2 March 2026, for the 37.9% / 31.2% / 31.0% split across 863,000 keyword result pages and 4 million AI Overview URLs, and for the note that the 76% to 38% comparison is not like-for-like.

Vishwakarma, Kumar and Jamidar, What Gets Cited: Competitive GEO in AI Answer Engines, Proceedings of the 49th International ACM SIGIR Conference, July 2026, for the 252,000 trials, the four unanimous gatekeepers, the two formatting factors that had no impact, the 2026-versus-2019 date definition, and the authors’ own statements of their testbed’s limits.

Seer Interactive, Study: Content Recency’s Impact on AI Visibility in 2026, 24 July 2026, for 72% updated within a year against 42% published within a year on a base of 4,124 cited pages, and for the manufactured-freshness sentence.

Google Search Central, Optimizing your website for generative AI features on Google Search, for the index and snippet eligibility requirement, the statement that Google Search ignores llms.txt and similar files, and the chunking line.

Ahrefs, We Analyzed 137K Sites: 97% of llms.txt Files Never Get Read, 15 June 2026, for the 137,210 domains, the 28% adoption figure that Ahrefs calls an upper bound, and the 97% of files that were never fetched.

Allmo, llms.txt, 23 January 2026, for one llms.txt page among 94,614 cited URLs.

Ahrefs, We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved., 11 May 2026, for the matched design, the three surface results, the 1 in 2,500 statement and the 12 daily citations figure.

Ahrefs, Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews, 12 December 2025, for the YouTube correlation of about 0.737, branded web mentions at 0.66 to 0.71, site pages at about 0.194, and the publisher’s own correlation disclaimer.

Surfer, Why Almost 40% of AI Citations Come from Your First 100 Words, updated 17 September 2026, for 38% of citations drawn from the first 100 words across more than 100,000 placements in more than 10,000 AI Overviews responses.

OpenAI, Overview of OpenAI Crawlers, for the independence of the user-agent settings and the consequence of opting out of OAI-SearchBot.

Zhao and Berman, Strategic Response of News Publishers to Generative AI, version 4, 15 April 2026, for the 7% post-blocking decline in weekly visits, the 24 blocking sites of which 19 added only GenAI directives, and the identification window.

SE Ranking, ChatGPT Shows Ads for 1 in 4 Commercial Prompts, 10 August 2026, for 50,006 commercial prompts, the 96.37% of placements with no advertiser citation, and the 0.09% exact-URL overlap; and Google AI Mode ads for the 11.53% and 1.95% overlaps and the matched comparison.

LLM Pulse, ChatGPT ads study, for 36,153 US answers and the 2.8% of ads pointing at a domain the same answer used as a source.