Generative Engine Optimization for small sites

The 30-second answer: generative engine optimization for a small site is five things, and none of them is a trick

Be fetchable as plain text by the crawlers that feed answer engines. Lead every page with the answer to one question, with the number and the date in the first paragraph. Give the assistant something it can cite: a statistic with a source, a quotation, a named reference. Keep every copy of each fact consistent, in structured data, in llms.txt, and across your old posts. Measure with logs and a monthly audit, not with rank-tracking tools built for a different problem.

Two reasons this is the whole list. Google’s guidance for AI features says there are no additional requirements beyond ordinary indexing, and the one controlled study of the question found that content changes a small site can make in an afternoon raised visibility by about a third, while the classic trick, keyword stuffing, did nothing.

Where the term comes from, and what was measured

The phrase was coined in a paper titled GEO: Generative Engine Optimization by Pranjal Aggarwal and colleagues, posted in November 2023 and published at KDD 2024. The authors built GEO-bench, a set of 10,000 queries from nine sources, and tested how editing a page changed how much of it appeared in a generated answer, using a position-adjusted word-count metric and a subjective impression score.

Three findings carry over directly to a small site. The methods that worked best were adding citations, adding quotations from credible sources, and adding statistics, each improving visibility by roughly 30 to 40 percent on the word-count metric; improving fluency and readability also helped. Keyword stuffing gave little or no benefit. And lower-ranked sites gained more from the same edits than top-ranked ones, which is the opposite of how classic search usually rewards effort.

The vendor side has its own picture. Google’s page on AI features and your website, last updated December 2025, lists its recommendations: allow crawling in robots.txt and in any CDN or hosting layer, make content findable through internal links, provide a good page experience, make important content available in textual form, support it with images and video where relevant, keep structured data matching the visible text, and keep Merchant Center and Business Profile information current. It adds that AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics, and that no new machine-readable files or markup are needed.

The five things

1. Be fetchable as text

Nothing else matters if the fetch fails. Allow the search crawlers (OAI-SearchBot, Claude-SearchBot, PerplexityBot) and Googlebot in robots.txt; check that your CDN or firewall does not challenge non-browser requests; and make sure the content is in the HTML as served. Google’s list puts crawling first and textual content fourth; the crawler comparison explains why the AI side cannot be assumed to render JavaScript. Test with curl, not a browser.

2. Lead with the answer to one question

Query fan-out means an assistant runs several specific searches, not one broad one. A page that answers one specific question in its first paragraph, with the number and a date, is what those searches retrieve and what the model can quote. Write the question as the heading, answer it in two or three sentences, then explain. A page that covers six questions loosely is retrieved for none of them.

3. Give the assistant something to cite

This is the finding from the paper. A sentence with a figure and its source, a quotation from a named person or document, a dated reference to a standard or a study: these are the parts of a page that survive into a generated answer. For a small business that means your own numbers (how many jobs, what turnaround, what the warranty covers), stated plainly, with the month they were true.

Before:
We offer fast, reliable bicycle repairs at competitive prices.

After:
As of October 2026 we complete most repairs in 2 working days; the shop
has serviced about 1,800 bikes since 2019. Frame repairs carry a 2-year
warranty (see the warranty page). Prices start at 35 EUR for a tune-up.

4. Keep every copy of a fact consistent

An assistant that finds two versions of a price, a plan name, or an opening time picks one. Google asks that structured data match the visible text; the same rule applies to llms.txt, old blog posts, and third-party listings you control. Put the dated summary in llms.txt, generate structured data from the same source as the page, and sweep old posts when a fact changes. The pricing article walks through the sweep for the most common case.

5. Measure with logs and a monthly audit

Three signals, none of which needs a paid tool: the search crawlers and user fetchers appearing in your server logs; referrals from AI domains in analytics; and a monthly audit where you ask three assistants your ten most important questions and record whether you are cited and from which page. The measuring article has the dashboard.

What you can skip

• Keyword stuffing and density targets. The one controlled study found little or no benefit. Write the answer; the words follow.
• Special AI markup or files as a requirement. Google says none is needed. llms.txt is useful for the readers that use it; it is not a ranking factor anywhere.
• Tools that promise AI rankings. There is no ranking to track. Visibility is per question, per assistant, and changes with the model. A spreadsheet of your ten questions does the job.
• Hundreds of generated pages. Fan-out rewards one good page per real question, not volume. A small site wins by being the clearest source on its few topics.
• Blocking training bots as a tactic. It is a legitimate policy choice, but it does not change citations; the search crawlers and user fetchers are different agents.
• Chasing the acronyms. GEO, AEO, AIO, and LLMO describe the same practice. The complete guide’s jargon section sorts them out; the work is identical.

An afternoon plan for a site under 100 pages

1. Fetch robots.txt and your three most important pages with curl. Fix anything blocked or missing from the HTML.
2. List the ten questions customers actually ask. Map each to one page; create a page for any that has none.
3. Rewrite the first paragraph of each of those pages to answer the question with a number and a date.
4. Add one citable fact per page: a statistic with its source, a quotation, or a dated reference.
5. Write llms.txt with a dated summary and those ten pages under named sections. Check that structured data, if any, says the same things.
6. Ask three assistants the ten questions. Record the answers and the cited pages. Repeat monthly.

Checking the first step

Step one is the one most sites fail without knowing it. A checker fetches your pages the way a bot does and reports what is reachable as text.

FAQ