Top generative engine optimization strategies for AI visibility in 2026: the evidence behind each one
The guides that rank for this question recommend roughly the same dozen or so strategies: write passages an assistant can lift, add statistics, mark the page up with schema, keep it fresh, earn mentions and links elsewhere, show who wrote it, let AI crawlers in. What stands behind those recommendations varies far more than the lists suggest. A few are documented by the platform that runs the answer; several rest on a study whose design is narrower than the advice built on it; a handful we measured ourselves; and for some we could find nothing at all. We read six strategy guides on 12 October 2026 and sorted every strategy they recommend into one of four classes of evidence:
- measured by us — a test this desk ran, with its sample and its limits;
- documented by the platform — the company running the search or the crawler says so in its own documentation;
- a study read on its own terms — published research, read for what it actually measured;
- nothing we could find — no test, no document, or a figure whose trail ends somewhere else.
One disclosure before the list. LiamVi builds a tool that records which sources AI assistants cite in each response, so the way anyone weighs a path into those answers touches our own business. Read the classes below as our view of what counts as evidence, not as a neutral standard. We are a small team and our own sample is small. An earlier version of this page, published on 8 September 2026, led with four tests of ours and never gave you the list; this version gives you the list first.
The order is not a ranking. An evidence class tells you how much is known about a strategy, not how well it works, and a strategy with nothing behind it may still help your site. Where a decision depends on your business, we say so and leave it with you.
Which strategies do the ranking guides recommend?
Six of the pages the search engine showed for this question are strategy guides, and between them they recommend the same core moves. Google's own guide to its AI features, last updated 10 July 2026, ranks first. Semrush's one-minute answer to this exact question, dated 11 September 2025, names six. HubSpot's eight best practices, by Ramona Sukhraj, was updated on 3 September 2026. The llmrefs guide sets out six numbered practices and carries no date. Profound's ten-step framework, by Josh Blyskal, is dated 1 July 2025, and Kepler Group's five practices are undated.
By our count, which is a reading and not a measurement, every one of the six asks for clearly structured content in some form. Four ask you to keep content fresh, four recommend schema markup and four recommend images or video. Two suggest a Wikipedia entry for your brand, and two suggest an llms.txt file. All six tell you to monitor your brand's visibility in AI search. The pages we read, and the date we read them, are listed under Sources.
None of the six says, item by item, what evidence stands behind each recommendation. That gap is the reason for this page. It is not a gap in the field as a whole: Olivier Martinez's critical survey of 45 studies, a single-author preprint posted on 15 July 2026, concludes that "already-retrieved content can causally alter its citation or use, but no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior." What follows is narrower than that survey: the tactics a reader meets on this one results page, each with the class of evidence we found behind it.
Which strategies did we measure ourselves?
Five of the generative engine optimization tactics on these lists, plus one that none of them names, sit inside our own harvest: 33 keywords from three neighbouring industries, taken in July 2026 and mostly on ChatGPT, with Gemini and Perplexity sampled far more thinly. That is small and early; the set gets larger each week as the sweeps run, and if a result moves, we publish the change.
Choosing a question an assistant will search for
In our test, phrasing was the difference between an answer with no citations and an answer with some, and none of the six lists it as a strategy. We wrote the prediction down first, then put five topics through four phrasings each, from a conceptual "what is X" to a dated question and a recommendation tied to a place. Not one of the five conceptual phrasings drew a citation; every one of the five dated phrasings did. The mean number of pages cited went from 0.0 at the conceptual end to 6.4 for the dated questions. Two topics moved against the run at the top of the ladder, one from six pages to two and another from twelve to one, and the replication on Gemini covered just two queries. Five topics, mostly on ChatGPT, is a small test, and each phrasing was recorded once, with no repeat to tell phrasing apart from the variation between runs of one prompt; the direction held, but the rates will not carry over to every assistant, and Perplexity in particular cites much more freely. Our research page prints the direction and its limits; the means and the two exceptions come from the same run's full record. On one small test, we read this as a commissioning decision for your next brief: frame your subject around a current state, a named thing or a comparison, and we think your odds of being looked up rise. They do not become a promise.
Writing passages that can be lifted
Passage selection is the one tactic on these lists our own evidence directly supports, though it shows that a sentence gets chosen, not what makes one choosable. HubSpot's first practice is to lead with a clear, direct answer, llmrefs asks you to structure content so an AI system can extract it, and Profound recommends TL;DR blocks. In our harvest some citations pointed at one sentence within a page, not at the page as a whole, and one article drew four citations, each linked to its own passage, as the research page records. For an AI Overview or AI Mode, the official documentation says there is no need to break content into small pieces.
Answering the questions people search for
Answering the People Also Ask questions came back null in our harvest, and it was our own hypothesis. The quoted pages covered 58.8% of those questions and the passed-over pages covered 58.2%, which is no difference worth the name; the research page publishes those two rates without the number of questions behind them, so read them as a direction, not a precise gap. Answering questions still makes an article better to read; it did not separate the quoted pages from the passed-over ones. The format is a separate matter from the content, and our read on FAQ sections sets out how little the published figures say about the section itself.
Domain authority and backlinks
Higher domain authority did not decide which page ChatGPT quoted: across 21 matched pairs, the page on the stronger domain came out ahead in 11 and behind in 10. Profound asks you to earn citations from high-authority domains, and llmrefs says backlinks remain part of the job. Our pairs were matched on keyword only, so content, age and links differ inside each pair too. By our calculation, 11 wins in 21 fit a true rate anywhere from about 32% to 72% (a 95% Wilson interval), so the sample rules out only an advantage that wins nearly every pair; a small, moderate or even sizeable one remains possible. Our sweeps also keep turning up young sites cited alongside very large ones. Links are a different object from authority, and the published link studies each measure their own outcome, as our look at backlinks and AI search shows.
Ranking well in traditional SEO
The overlap between ranking and being quoted was small in our harvest: of the 73 pages ChatGPT cited, 9 also sat in the top ten organic results for the same query, 12.3% of the citations we collected. Several of the lists say strong SEO feeds AI visibility, and Google says the same about its own AI features. That figure measures how far two sets overlap, counted in pages rather than domains. It is not a citation rate by rank, so it says nothing about whether ranking well makes a citation less likely, and it does not mean traditional SEO stops mattering.
Letting AI crawlers in
Crawler access is partly settled by the vendors' own documents and only lightly by our data. Semrush and llmrefs both put it near the top. OpenAI's crawler documentation separates two jobs: GPTBot "is used to crawl content that may be used in training our generative AI foundation models", while "Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers". The crawler list, last updated 14 July 2026, says Google-Extended "does not impact a site's inclusion in Google Search" and is not a ranking signal there. Our own measurement is early and has not been published on our research page yet, so there is nothing to link: blocking an AI crawler was about as common among the domains that were quoted as among those passed over, and only 1 of the 230 domains we checked blocked OAI-SearchBot, which is too little variation to test. Look at which user agent your robots.txt actually blocks before you decide why your site is or is not quoted. Server-side rendering has no published test we found behind it; Google says its own systems can process JavaScript as long as it isn't blocked, which is a statement about one crawler, not every crawler.
Which strategies rest on a study you can read?
Seven of the recommended tactics lean on published research, and in each case the research measured something narrower than the advice. We have read each one in full elsewhere; here it is one line and one link per strategy.
Adding statistics and quotations
The evidence for adding statistics is one benchmark study, and its measure is a share of an AI-generated answer, not a citation by a live assistant. Semrush, llmrefs and HubSpot all recommend it, and the study behind it is the 2023 paper that named the field, GEO: Generative Engine Optimization. Its nine methods are all rewrites of a page's text, tested on 10,000 queries that its authors describe as "repurposed for generative engines, along with synthetically generated queries"; gpt-3.5-turbo wrote each response from the top five search results. The paper's "up to 40%" is a page's share of that generated answer, and it tested no lists. Keyword stuffing is one of the methods the paper itself files under "Non-Performing", the paper's own label for the methods that offered little or no improvement. Our reading of what the paper measured covers which pages gained and what later tests found.
Schema markup
The largest controlled test of schema markup, published by Ahrefs on 11 May 2026, was run on pages that an AI system was already citing heavily. Four of the six recommend it, while Google's documentation says structured data is not required for its AI features. That leaves open the case most readers care about, a page nobody cites yet, and our read on schema markup explains why that boundary matters.
Keeping content fresh
The freshness datasets, Ahrefs's July 2025 analysis of the URLs AI assistants cite among them, describe how old the pages AI already cites are, which is a different thing from what happens when you update one. Semrush, HubSpot, llmrefs and Profound all recommend regular updates, and llmrefs gives its own figure "from our data" with no sample printed. We found no published test of a refresh, as our read on content freshness sets out.
Earning mentions across the web
The figure most often quoted for earning coverage elsewhere, from Muck Rack's May 2026 What Is AI Reading? report, is a share of citations across a wide category of sources, and it records which pages were chosen, not why. Semrush says an unlinked brand mention can boost your AI visibility, and llmrefs suggests getting your brand named in the pages AI already cites, so that a mention sits where the assistant already reads. Our read on earned media separates the press coverage from the rest of that category. Google's guide adds a line worth keeping beside this one: seeking inauthentic mentions "isn't as helpful as it might seem." Placing your own material on other sites in the hope an assistant reads it has a name of its own, LLM seeding.
Comparison and "best X" articles
The published figures on comparison articles, from Seer Interactive's tracking and other studies, are shares of citations, which cannot say whether any one list is likelier to be picked. Profound tells you to prioritize comparative articles; our read on listicles explains the difference between a share and a rate.
Publishing original research
The one published comparison we found that also counted the pages AI skipped, Eric Lancheres's July 2026 test for On-Page.ai, found no originality advantage for the pages AI Overviews cited. Profound and Kepler both point to original, data-driven content; our read on original data sets that comparison beside the figures that count only pages already cited.
Author bios and expertise signals
A published byline test, Seer Interactive's byline case study, found no across-the-board lift in citations. HubSpot, llmrefs and Profound all ask for author credentials and E-E-A-T signals, and HubSpot's related advice, being specific about the people and companies you mention, is entity clarity, for which we found no published test; both appear in our audit of an eight-signal checklist.
Which strategies does a platform document?
Five tactics are backed by the platform's own words, and in every case the platform is describing its own surface. Google's documents describe AI Overviews and AI Mode and say nothing about how ChatGPT or Claude picks a source; a document is a statement of design, not a measurement.
Non-commodity content
Original, useful content is the one recommendation the search engine puts above every other in its documentation. It says content that is "unique, compelling, and useful" will likely influence presence in its AI features "more than any of the other suggestions in this guide", and tells you: "Don't just recycle what others on the internet have already said, or could easily be produced by a generative AI model."
Images and video
Images and video are documented as a way to show up in AI Overviews and AI Mode, whose documentation says the features "can bring in relevant images and video". HubSpot attaches a figure to this practice, crediting "Research from Princeton and Georgia Tech" with 40% more AI citations for content with images, charts and videos. The link on that sentence opens a 2024 statistics paper on stratified randomization, and the GEO paper, whose authors include Princeton researchers, tested no images. We could not find the study the sentence describes; a link can point at the wrong address, and the figure may have an owner we did not reach.
Fan-out sub-queries
Query fan-out is documented, and so is the line you should not cross with it. llmrefs advises ranking for the shorter searches an assistant generates from a user's longer question, each one a keyword in its own right. The search engine describes that same fan-out, and says that creating separate content for every variation primarily to manipulate rankings or an AI response "violates Google's scaled content abuse spam policy".
Writing for people
No special style is required for AI search, in Google's own words. HubSpot's seventh practice is to write like a real person, and the guide agrees from the other side: "You don't need to write in a specific way just for generative AI search."
Monitoring your visibility
Monitoring is how you find out where you stand, not a strategy that changes it, though all six recommend it in some form. For AI Overviews and AI Mode, their documentation points to a Generative AI performance report in Search Console. Profound suggests auditing how each AI platform describes your brand and tracking 20 to 30 prompts per core topic, run once a day, to monitor visibility over time; llmrefs calls share of voice the most important metric, and for a team without a tool it describes a manual check: ask 10 to 20 queries, note whether your brand appears and which sources each response cites, and repeat monthly. Those are ways to track your AI visibility, not evidence that tracking moves anything. One caution applies to any report or tracking tool you use: the same prompt can return a different AI response from one run to the next, so a single check is one sample.
Which strategies have no evidence we could find?
Four tactics on these lists, or sold beside them, have nothing behind them that we could trace. "Nothing we could find" is a report of one desk's search, made on 12 October 2026, not proof that no evidence exists anywhere.
Tables and other structured formats, recommended by Profound and llmrefs, most often travel with a 2.5x figure whose trail ends at a page that does not contain it, as our read on tables shows. A Wikipedia entry for your company, suggested by Semrush and llmrefs, can reach an answer through training data or through live search, and the figures usually quoted describe the encyclopedia as a whole rather than one company's article; our read on Wikipedia separates the two routes. Content syndication is on none of the six lists but is sold elsewhere as a route to citations, and the word covers three different practices; our read on syndication says which of them has been measured and by whom. An llms.txt file, offered by Profound as a bonus and by llmrefs as worth considering, is ignored by Google Search on its owner's word, and no test of it on another AI platform turned up in our search.
Which GEO tactics does Google say you can ignore?
Five tactics can be skipped, by the search engine's own account, and the list is short enough to give in full. On special files: "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." On chunking: "There's no requirement to break your content into tiny pieces for AI to better understand it", and "There's no ideal page length". On rewriting: "You don't need to write in a specific way just for generative AI search." On mentions: "seeking inauthentic 'mentions' across the web isn't as helpful as it might seem." On structured data: "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add", though it is still recommended for rich results. Its summary is "Prioritize effective SEO strategies over 'AEO/GEO hacks'".
The documentation also calls it "completely fine" to keep an llms.txt file for other services that use one. All of this describes one company's surface. We tested none of the five; this is documentation quoted, not confirmed.
The one requirement it does set is a floor: a page has to be in the index and able to appear with a snippet, and the site must also be included in those features in Search Console. Meeting every requirement guarantees nothing: "Indexing and serving aren't guaranteed."
Is generative engine optimization different from SEO?
For its own AI features, Google says no: optimizing for its generative AI search "is optimizing for the search experience, and thus still SEO." On the assistant we measured most, the answer looked different, because the pages the search engine ranked and the pages the assistant quoted had very few members in common in our harvest. Both statements hold, since they answer separate questions: one describes how that company built its AI features on its existing ranking systems, and the other describes which pages a separate assistant chose. Traditional SEO keeps doing its job, technical SEO included, and generative engine optimization sits beside it: the work of earning a place in an AI response, measured another way. If your SEO is sound, much of what the six lists recommend is already part of it.
How do you check a GEO claim — including ours?
A tested claim and an asserted one can be told apart with three questions, and they apply to this page as much as to any list of tactics. The search engine's guidance on third-party SEO tools and advice, last updated 5 June 2026, says "Third-party tools don't have access to our internal ranking data. They can't guarantee performance." That describes us: we see inside no AI system, and no citation is ours to promise.
What was measured, on how many of what? Without a sample, a comparison group and a date, a recommendation is an opinion presented as a result. Ours is 33 keywords in three industries, mostly ChatGPT, July 2026.
What came back null, and was it published? Real testing produces failures, since most ideas do not survive the data. A provider who lists only what works either has not run the tests or is not sharing them; our own nulls sit in the section above, beside the one test that moved.
Is it about the surface you care about? An AI Overview, ChatGPT search, Gemini, Perplexity and Claude choose sources in their own ways, so a result on one is not a result on the others. A user asking the same prompt twice can get two sets of sources.
One SEO writing habit we advise against on judgement, not on a measured result: we think a sentence that says "according to a report, X" hands the quotable line to the report, so state the fact in your own words and keep the link on it. And ask any tool, ours included, how its score was checked: the first version of the draft score our own tool gives rewarded matching the pages that already rank, and once we found that, we turned it into a list of what an article is missing rather than a grade of how good it is.
What people ask about GEO strategies
How do you do generative engine optimization in practice? Start with the decision that comes before writing, which is whether the question your page answers is one an assistant looks up at all; then write passages that make sense when lifted out alone, each carrying its own figure and source. Weigh everything else on the lists against the evidence class beside it above.
What counts as a GEO best practice? A best practice worth the name says what was tested, on what sample and when, and a strategy without those three is a suggestion. Of the tactics on the six lists we read, a handful have platform documentation behind them, several rest on studies narrower than the advice, and some have nothing we could trace.
What does a GEO example look like? In our harvest, one article earned four citations, and every one of them linked to a separate passage inside it. The page did not change between those four selections; four of its sentences each answered something.
Is a GEO course, tool or service worth paying for? We have not measured any course, nor any agency's generative engine optimization service, so there is no result of ours to report. Take the three questions above to whoever is selling, and weigh the answers against what they would cost your business; that judgement is yours.
What this ledger cannot tell you
An evidence class is not an effect size, and a practice with nothing behind it is untested, not disproved. Google documents its own surface, and OpenAI documents its crawlers; neither measures every assistant. The ranking set we read is the one captured for this search on 3 September 2026, reopened on 12 October 2026, and pages like these change. Our own measurements are mostly ChatGPT and one harvest at one moment, and we cannot yet show that adding any property causes a citation; our weekly tracking is gathering the outcome data that would test it. If a strategy here moves to a better class of evidence, we will move it.
Sources
- Optimizing your website for generative AI features on Google Search — Search Central guide, last updated 10 July 2026; checked 12 October 2026.
- Common crawlers — crawling infrastructure documentation (moved from Search Central), last updated 14 July 2026; checked 12 October 2026.
- Guidance on third-party SEO tools, services and advice — Search Central, last updated 5 June 2026; checked 12 October 2026.
- OpenAI — Bots and crawlers — checked 12 October 2026.
- Semrush — What are the top generative engine optimization strategies for AI visibility? — 11 September 2025; checked 12 October 2026.
- HubSpot — 8 generative engine optimization best practices — updated 3 September 2026; checked 12 October 2026.
- llmrefs — GEO: The 2026 Guide — undated; checked 12 October 2026.
- Profound — 10-step GEO framework — 1 July 2025; checked 12 October 2026.
- Kepler Group — GEO: The Future of Digital Visibility — undated; checked 12 October 2026.
- Aggarwal et al. — the GEO paper, arXiv 2311.09735 — November 2023, KDD 2024; checked 12 October 2026.
- Martinez — a critical survey of the field (2023–2026), arXiv 2607.14035 — 15 July 2026; checked 12 October 2026.
- Ahrefs — schema and AI citations — 11 May 2026; checked 13 October 2026.
- Ahrefs — fresher content and AI assistants — 28 July 2025; checked 13 October 2026.
- Muck Rack — What Is AI Reading? — May 2026; checked 13 October 2026.
- Seer Interactive — listicles in AI search — 25 February 2026; checked 13 October 2026.
- On-Page.ai — do AI assistants cite original content? — 15 July 2026; checked 13 October 2026.
- Seer Interactive — author bylines and AI visibility — undated; checked 13 October 2026.
- Our research — methods, samples and limits — the framing test, the passage evidence, the matched pairs, the ranking overlap and the question-coverage null; last updated 12 August 2026; checked 12 October 2026.
The ranking set read. Captured in the United States on 3 September 2026, in the order returned; every page reopened on 12 October 2026, with the dates each page prints.
- developers.google.com — the search engine's own guide to its AI features; last updated 10 July 2026, the same date as in September.
- tryprofound.com — a ten-step framework; dated 1 July 2025.
- semrush.com — Generative Engine Optimization: A Practical Guide, not the one-minute answer; published 6 June 2025, updated 16 April 2026.
- beomniscient.com — a comparison of tools, not a strategy guide.
- keplergrp.com — five practices; no date printed.
- blog.hubspot.com — eight best practices; updated 3 September 2026.
- llmrefs.com — a guide with six practices; no date printed.
- truedigitalcom.com — a post on AI visibility; the site refused our requests, so nothing here draws on it.
- seranking.com — a product page, not a strategy guide.
Not in that capture: Semrush's one-minute answer (11 September 2025), found in the live results on 12 October 2026, the sixth guide read.
