Generative engine optimisation: how expert businesses get cited by AI search

Generative engine optimisation (GEO) is the work of making your content easy for AI search to find, understand and cite. For an expert business it is mostly ordinary good search practice: pages that answer the question first, named authors, original evidence and HTML that crawlers can read. Nobody, including Google, can guarantee that an AI answer will cite you.

What do GEO and AEO mean?

GEO is generative engine optimisation. AEO is answer engine optimisation. Both labels describe work aimed at appearing inside AI-written answers, such as Google's AI Overviews and AI Mode or the answers in ChatGPT search.

The research paper that introduced GEO, by Pranjal Aggarwal and colleagues, was first posted to arXiv in November 2023 and accepted to the KDD 2024 conference (arXiv (opens in a new tab)^[12]). Google treats the labels as ordinary SEO. Its guide, last updated on 10 July 2026, puts it this way: "From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO" (Google Search Central (opens in a new tab)^[2]).

How do AI answers choose their sources?

The engines described in Google's guide and in the GEO paper both retrieve candidate pages and write the answer from them. A page that can't be retrieved can't be quoted, however well it is written.

Google describes its version in its guide. Retrieval-augmented generation (RAG) relies on Google's core ranking systems "to retrieve relevant, up-to-date web pages" from the Search index. The system then reviews those pages and writes a response with links to the ones that support it. Query fan-out is "a set of concurrent, related queries generated by the model" to fetch more results. Google's example is a question about a weedy lawn, which might also trigger searches on herbicides and prevention (Google Search Central (opens in a new tab)^[2]).

The GEO paper models an engine as query reformulation, retrieval of ranked sources, summarisation and a final response with citations (arXiv (opens in a new tab)^[13]). Google says AI Overviews and AI Mode "may use different models and techniques, so the set of responses and links they show will vary" (Google Search Central (opens in a new tab)^[1]). OpenAI's crawler documentation says sites that opt out of OAI-SearchBot "will not be shown in ChatGPT search answers", though they "can still appear as navigational links" (OpenAI (opens in a new tab)^[14]).

None of these pages publishes a formula for choosing between eligible pages. Treat any list of "AI ranking factors" as inference from testing, not documentation.

What does Google say about its AI features?

The AI features page on Google Search Central states it plainly: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary" (Google Search Central (opens in a new tab)^[1], last updated 10 December 2025). To appear as a supporting link, a page "must be indexed and eligible to be shown in Google Search with a snippet". Meeting every requirement still guarantees nothing. The same page says, "Indexing and serving isn't guaranteed."

On 15 May 2026 Google added a dedicated guide, "Optimizing your website for generative AI features on Google Search" (Search documentation updates (opens in a new tab)^[3]). Its central advice concerns content: unique, useful pages will likely shape a site's long-run presence in generative AI search more than any other suggestion in it.

Four dated entries from 2026 matter here.

The dates come from Google's updates log (opens in a new tab)^[3] and the Search Console announcement of 3 June 2026 (opens in a new tab)^[6].

Search Console also has a Search generative AI control under Settings. Including a site is the default. Excluding it removes the site's links and content from AI Overviews, AI Mode and the generative AI features in Discover. Google says the control reached all websites worldwide as of 31 August 2026 (Search Console Help (opens in a new tab)^[8]). Its generative AI guide lists inclusion in this control as a condition of eligibility, on top of the technical requirements.

What did the GEO paper test, and what did it find?

The study asked whether rewriting a page changes how much of an AI answer is credited to it. The authors built GEO-bench, 10,000 queries drawn from nine sources across 25 domains, and ran its 1,000-query test split through a simulated engine: the top five Google results for each query, then gpt-3.5-turbo writing a cited answer. For each query, one source page was picked at random and rewritten by a language model using one of nine methods. The paper says its scores are averages over five random seeds, and the table below is its Table 1 (Aggarwal et al., KDD 2024 (opens in a new tab)^[13]).

Two metrics were used. Position-Adjusted Word Count counts the words in sentences that cite the page, weights earlier sentences more heavily and divides by the answer's length. Subjective Impression averages seven GPT-3.5 ratings, such as relevance and likelihood of a click. Both use the paper's own scale, where the unmodified baseline scores 19.3 on each in Table 1 and higher is better.

Quotation addition was the strongest method on both metrics. Against the 19.3 baseline, its 27.2 is the 41% improvement the authors report on the first metric, and its 24.7 is a 28% improvement on the second. Statistics addition and cited sources also beat the baseline on both. Keyword stuffing, the classic search tactic, split by metric: 17.7 on Position-Adjusted Word Count, below the 19.3 baseline, but 20.2 on Subjective Impression, above it. The authors conclude that it offers "little to no improvement". They also report no significant gain from a persuasive tone, although the table shows a smaller lift, so treat that row with care.

The paper is not fully consistent. Its appendix says Table 6 holds the five-seed results with standard deviations, but Table 6 lists a baseline of 19.8, not 19.3. It also shows keyword stuffing level with that baseline on the first metric (19.8, and 20.6 on the second) and quotation addition at 27.1, about 37% above its baseline by this article's arithmetic, against the 41% in Table 1's caption. The paper does not explain the gap. The direction agrees, but cite Table 1 as Table 1, and do not read Table 6 as independent confirmation of its percentages.

The authors repeated the test on Perplexity with 200 test queries, uploading each page's text as a file (Table 5). Quotation addition scored 29.1 against a 24.1 baseline on the first metric. Keyword stuffing scored 21.9 on the first metric, below baseline, and 28.1 against 24.7 on the second, above it.

What the study cannot tell you:

• Business outcomes. It measures a page's share of a simulated answer. No clicks, enquiries or sales were counted, and the click-likelihood score is a GPT-3.5 rating, not observed behaviour.
• How today's engines behave. The engines were a purpose-built pipeline on gpt-3.5-turbo and Perplexity fed uploaded files, so its own retrieval went untested. The authors write that methods "may need to adapt over time as GEs evolve", and they did not evaluate effects on search rankings.
• Whether the additions were accurate. A language model produced the added quotations, statistics and citations, and the limitations section does not address their accuracy. The result describes model-added material, not verified evidence. On a real page, evidence must be genuine and checkable, which is an editorial rule rather than a finding of the study.
• Anything after 2024, or any Google feature. Google's 2026 guide does not mention the paper, publishes no test data and does not say the result transfers to AI Overviews or AI Mode.

What should an expert business do?

Start with the pages a buyer reads before they get in touch. Each practice below rests on a documented source or is marked as editorial judgement.

Answer first. Put the direct answer in the opening lines, under a heading phrased as the question, then give the evidence and the limits. Google says "You don't need to write in a specific way just for generative AI search", so the reason to lead with the answer is the reader, who wants it first. Answer-first writing for people and AI covers the format.

Name the author and the business. Google's helpful-content guidance asks whether it is "self-evident to your visitors who authored your content" (Google Search Central (opens in a new tab)^[9]). Give each page a byline that links to an author page with real credentials, and spell business, people and service names the same way across your site and profiles. That second habit is editorial judgement, not a documented ranking factor; entity SEO for experts sets out the reasoning. Google calls fabricated creator profiles, including made-up names and false credentials, "a form of deception".

Publish what only you can. Google's generative AI guide asks for "a unique point of view" and warns against content that "could easily be produced by a generative AI model" (Google Search Central (opens in a new tab)^[2]). For an expert that means your method, your reasoning on real situations (anonymised, with permission) and measurements from your own work, each with a date and a note on how it was gathered.

Cite accurately. In the GEO study, model-added citations and quotations helped, and the study did not check whether they were true. The editorial rule here is stricter: link the primary source, quote it verbatim and briefly, and give the year, the population and the measure for every figure. Never add a number you can't trace.

Keep pages crawlable. Google's AI features page lists crawling allowed in robots.txt and by any CDN or hosting layer, plus important content in text form (Google Search Central (opens in a new tab)^[1]). OpenAI treats OAI-SearchBot (ChatGPT search) and GPTBot (model training) as independent robots.txt settings, and says a robots.txt change can take about 24 hours to reach ChatGPT search (OpenAI (opens in a new tab)^[14]). Google separates Google-Extended, a robots.txt control for Gemini training and grounding, from its Search crawling, and says it does not affect inclusion in Google Search (Google Search Central (opens in a new tab)^[5]). How AI crawlers read your website goes through each operator's controls and checks.

Match markup and dates to the visible page. Google's AI features page asks for structured data that "matches the visible text on the page" and says no special schema.org markup is needed. For dates, Google recommends a visible label such as "Last updated", matching datePublished and dateModified in a CreativeWork subtype such as Article, and no future dates. Its systems estimate when a page was "published or significantly updated", and the dates must describe the page's own publication or update (Google Search Central (opens in a new tab)^[10]). As editorial judgement, move the date only for a material change, never to look fresh. Skip FAQ markup if a rich result is the only aim.

A worked example (hypothetical)

Take a fictional consultancy whose "How we work" page says only that it takes "a collaborative, tailored approach" to each client. A buyer who asks an AI assistant what an onboarding audit covers gives that page nothing to quote. Rebuilt around the question, the same page changes like this. Square brackets mark placeholders that the firm's own approved material must fill.

Nothing here is a result or a customer story. The point is the editorial decision: replace a claim about style with the facts a buyer's question needs.

Repeated questions from prospects are the best raw material for these pages.

What does not work?

Six common tactics, and what the sources say about each. The table draws on Google's generative AI guide (opens in a new tab)^[2], the GEO study and Google's advice on outside services.

Plandora publishes an llms.txt file and an XML sitemap. That states what exists, not what it has earned, and this article claims no citations from either. Whether other AI products read llms.txt is covered in does llms.txt matter.

How do you measure AI search visibility?

Three kinds of evidence answer three different questions, and a rise in one does not show a rise in another. Crawler logs show access. A request from OAI-SearchBot or GPTBot means a bot fetched a URL, not that anyone saw the page or that an answer cited it. Referrals show visits. Analytics count people who arrived from a link, and miss answers that cited you without a click. AI search referrals: what the studies measure explains what published referral studies counted and what they left out. Saved answers show observed citations, and they need a procedure.

In Search Console, the Generative AI performance report counts impressions: how many times links to your site were shown in AI Overviews and AI Mode, by page, country, device and date (Search Console Help (opens in a new tab)^[7]). Read it as a measure of appearance, not of visits. Google says this data is also included in the overall Performance report, so treat appearances from the Generative AI report and clicks from the Performance report as different measures. Google says the report usually doesn't appear when a site has too few impressions in AI features, or may be excluded.

For the saved answers, follow the same steps each round:

1. Fix a set of 10 to 20 real customer questions, worded as customers word them, and keep the wording unchanged between rounds.
2. Ask each question in a fresh conversation, so earlier chats do not shape the answer.
3. Record the engine and mode, the date, the location and language settings, and the exact prompt.
4. Repeat each question several times per round. Answers vary: Google says the responses and links in AI Overviews and AI Mode will vary, and the GEO authors sampled five responses per query "to reduce statistical deviations".
5. Save the full answer text and every cited link, then note which of your pages, if any, were cited.

Treat before-and-after movement as observational. Engines, prompts, locations and your own pages all change at once, so a rise after an edit does not show that the edit caused it, and a single answer is never a result. Ten to twenty questions is a starting sample, not a picture of the whole market. Questions visitors put to your own Digital Mind (see Plandora's page for expert businesses) are a ready source for the list, and Insights can show top questions and answer coverage, depending on your plan.

What is a seven-step GEO checklist?

1. Confirm eligibility. Pages are indexable, crawling isn't blocked in robots.txt or by your CDN, important content is text, and the Search generative AI control includes your site.
2. Fix the question set and take a baseline. List 10 to 20 real customer questions, run them as described above, and save the answers before changing anything.
3. Rewrite priority pages answer-first. Follow the worked example: the direct answer, then scope, exclusions, evidence and limits.
4. Add original evidence you can trace. For a hypothetical consultancy, that could be its audit checklist with the date and how it was compiled, or an anonymised before-and-after of a client page shared with permission. Leave out any figure you can't source.
5. Show who wrote it and when. Add a real byline, an author page, an About page, consistent names and true dates.
6. Match structured data to the visible page. Markup should mirror the text a reader sees.
7. Repeat the measurement monthly. Rerun the same questions, keep crawler logs, referrals and saved answers separate, and record what changed. No step guarantees a citation.