6 min read
Generative engine optimization is the work of getting a business named, quoted, and linked inside an AI-generated answer. It overlaps heavily with SEO. It is not the same job, and the difference is worth being precise about.
Most of what gets written about GEO is either a rebranding of on-page SEO or a list of tactics with no mechanism behind them. This is an attempt at the mechanism: what actually determines whether a model cites you, in the order the inputs matter.
The one structural difference
A search results page is a list you choose from. An AI answer is a sentence you accept.
When the output is a list, being sixth still gets you seen. Some fraction of people scroll. When the output is a sentence, being sixth means you were not mentioned at all — there is no sixth position in a paragraph. Visibility stops being a gradient and becomes binary.
That is the whole reason GEO exists as a separate discipline. Everything else follows from it.
How an answer gets built
Simplified, but accurate enough to plan against. A generative search system does four things:
- Fans the query out. One question becomes several sub-queries. “Best physical therapist in Lincoln Park” might become queries about credentials, insurance acceptance, reviews, specialties, and hours. Google has described this fan-out behaviour explicitly.
- Retrieves candidates for each sub-query — from the search index, from its training data, and increasingly from live fetches.
- Selects sources it can use: pages it can parse, statements it can attribute, entities it can resolve confidently.
- Synthesises one answer and attaches citations to it.
You lose at step three far more often than at step two. Most businesses are retrievable. Fewer are usable — parseable, attributable, resolvable.
Why top-ten rankings stopped being a proxy
Ahrefs looked at 863,000 keywords and around four million AI Overview URLs in March 2026. Of cited pages, 38% ranked in Google’s top ten. 31.2% ranked 11 to 100. 31% ranked beyond position 100 or did not rank at all. The equivalent top-ten figure in July 2025 was 76%.
Ahrefs notes that part of that fall reflects better parsing in their own methodology rather than a pure behavioural change at Google, so treat 76-to-38 as a direction rather than a measurement. The direction is still the point: query fan-out pulls in sources that would never have surfaced for the original query.
Source: Ahrefs, March 2026, via Search Engine Journal.
The practical consequence: optimizing a page for one head term optimizes it for one branch of a fan-out. Coverage of the surrounding questions is what gets you into the other branches.
Ranking is no longer the mechanism. In July 2025, 76% of AI Overview citations came from top-ten pages. By March 2026 that figure was 38%. The model is reaching further down the index, and further outside it, than a rank tracker can see.
The five inputs, in order of leverage
1. Entity resolution
Before a model can recommend you, it has to be confident you are one thing. That confidence comes from agreement across sources: name, address, phone, category, hours, service area, the people attached to the business. Every disagreement is a small subtraction from confidence.
This is unglamorous and it is the highest-leverage work available to most local businesses, because almost nobody does it completely. Practically: pick one canonical string for every field and enforce it everywhere, including punctuation and suite numbers.
2. Corroboration from sources that are not you
Self-description is weak evidence. What third parties say is strong evidence. Ranked roughly by weight: established local and trade press, industry bodies and licensing registries, well-maintained directories, review platforms with genuine volume, and credible roundups.
The tactic that follows is not link building in the 2015 sense. It is working out which specific sources already appear in the citations for your category and city — you can read them straight off the AI answers — and earning a place in those.
3. Machine-readable structure
Schema markup is how you stop a system from having to infer. At minimum: Organization or LocalBusiness with complete contact and area data, Service for each thing you sell, FAQPage where you have real questions and answers, BreadcrumbList, and Person for anyone whose reputation carries weight.
Two rules that get broken constantly. Never mark up something that is not visible on the page. And if you are running an SEO plugin alongside a theme that emits its own Organization schema, turn one of them off — duplicate, conflicting entity definitions are worse than none.
4. Extractable answers
A page that gets cited usually contains a passage that stands on its own: a direct answer, in plain language, near the top, with the specifics in text.
- Lead with the answer, then explain. Not the reverse.
- Put numbers, prices, ranges, timelines, and exclusions in text — not in an image, a PDF, or a JavaScript-rendered widget.
- Use headings that are the questions people actually ask, phrased the way they ask them.
- Say what you do not do. Negative statements are unusually quotable, and they filter out bad-fit enquiries at the same time.
5. Freshness and maintenance
Stale pages get displaced. A page with a visible last-updated date, accurate current pricing, and no dead links is easier to trust and easier to quote. This is maintenance rather than strategy, and it is the input most likely to quietly decay.
What GEO does not do
It does not produce guaranteed placement. It does not work uniformly across ChatGPT, Gemini, Perplexity, Copilot and Google AI Mode — they weight sources differently and change without announcement. It does not replace classic SEO, because most of the retrieval still runs through a search index.
And it is slower than paid channels. Entity consistency shows up in weeks. Corroboration takes months. Anyone describing this as a fast win is describing something else.
How to measure it without fooling yourself
AI answers are non-deterministic. The same prompt returns different text at different times and to different users. A single check proves nothing, so measure like this:
- Fix a set of 10 to 20 prompts that mirror real buying questions for your category and city. Do not change them.
- Run them on a fixed schedule, in clean sessions, across every surface you care about.
- Log three things per run: whether you were named, who else was, and which sources were cited.
- Report the trend, not the snapshot. Share of answers over time is the number that means something.
The citation log is the most actionable artefact in the whole exercise. It is a ranked list of the sources you need to be on, generated by the systems you are trying to influence.
Related: how to find out if AI recommends your business — the ten-minute version of the diagnostic. Or have it run for you with the free AI Visibility Snapshot.
Found First works ranking, citation, and recommendation as one program. Scope and pricing are on what we do — published, no call needed to see a rate.