Generative engine optimization is the work of making a page the material an answer engine reads, trusts and quotes. That is the whole idea. It is not a new channel, and it is not a set of tricks that persuades ChatGPT to recommend your company. It is a measurement problem with a small number of observable outputs, and most of the noise around the term comes from people skipping the measurement.
We run a research site for exactly this reason. Our own tests are the only reason we’re willing to write about the subject, because the honest version has an uncomfortable finding in it: being cited by an AI answer and being recommended by one are two different outcomes. We’ve measured both.
What “generative engine optimization” describes in practice
Three surfaces are usually bundled under the label. Google’s AI Overviews appear above or inside a results page. Standalone assistants — ChatGPT, Gemini, Claude, Perplexity — answer in a chat window. Google’s AI Mode turns the results page itself into a conversation. They share a dependency: all of them need to read a page, and to read it they have to find it first.
That is where GEO meets SEO. The crawl, the index, the canonical tag, the rendered HTML, the hreflang pair — all of the unglamorous plumbing still decides whether an answer engine can see the page at all. A page Google cannot index is a page no answer engine can quote. If the indexing base is broken, no amount of answer-shaped formatting helps.
People shorten “generative engine optimization” to “geo seo” in conversation, and vendors use the two interchangeably. Treat them as one job with two names, and judge the vendor on what they measure rather than on which acronym they print.
Three counts that replace “AI visibility” as a feeling
“Are we visible in AI?” is not measurable as written. Split it into three counts that are.
- Coverage. For a given query, does an AI answer appear at all? This is a property of the query and the market, not of your site.
- Citation. When the answer appears, is your domain listed among its sources?
- Recommendation. Does the answer name your company as a suggested option?
Those three move independently. Coverage can be 90% while your citation count is zero, because the answer quotes other sites. A citation can exist with no recommendation, which is what happened to us. A recommendation can appear without a citation, because the model recalled the brand from somewhere else. If a report collapses them into one “AI score”, you can’t tell which of the three you’re buying.
AI Overviews, assistants and AI Mode are not one surface
They differ in ways that change the work. An AI Overview is generated inside a Google results page and draws on Google’s index, so its sources are pages Google has crawled and ranked. It is the surface most directly connected to classic SEO, and the one you can influence by fixing indexing and formatting.
A standalone assistant answers from a model plus whatever it can retrieve, and it may rely on browsing, on a search partnership, or on training memory. A brand can be named because it was read in a crawl this week or because it was absorbed into the model a year ago. Those two routes need different evidence, and a chat answer usually doesn’t tell you which one fired.
AI Mode blends the two: a conversational layer over Google’s index. In our panels the same prompt can return a sourced answer one day and a straight list the next. Report coverage per surface rather than averaged across all three, or the numbers will look random when they aren’t.
How we measure it, so you can copy the method
The panel is made of real questions, not keywords. A purchasing manager types “which supplier makes waterproof connectors for outdoor use” far more often in conversation than the head term a keyword tool suggests. We write 50 to 100 prompts per market, grouped by topic cluster and by buying stage.
The environment is clean and fixed. Logged out, one region, one interface language, the same day for the whole panel. We record four fields per prompt: whether an AI answer appeared, which domains it cited, which companies it named, and in what role our brand appeared. Raw answers are saved, because annotation rules change and you’ll want to re-read the text.
Then we re-run the same panel monthly. AI answers fluctuate, so a single run is a point measurement, not a trend. Reporting one snapshot as “our AI visibility” is the most common error we see in vendor decks.
What our own measurement showed
Two results set our expectations for every client conversation.
First, coverage is already large in markets people assume are behind. In September 2026 we tested 500 Chinese-language B2B queries and 80.9% returned an AI Overview. Coverage is highest on conceptual terms and lowest on technical, operational lookups. For a site without authority, that bottom range is the opening, because classic organic results still take the click there. The method and the query list are in our AI Overview coverage study.
Second, citation is not recommendation. We ran 72 questions in a clean environment. The answers named 18 companies in total. Our own sites were cited three times — and each time the citation came from a “how to choose a service provider” list rather than from a recommendation of us as the provider. Read plainly: the answer engine used our page as a checklist source, then suggested other companies. That’s a real outcome, and it’s the one most GEO pitches don’t mention.
Why “cited but not recommended” is the normal case
Answer engines behave differently when they describe a market and when they endorse a vendor. For a comparison or a method, they lean on neutral, specific, well-structured pages. For a recommendation, they weigh brand familiarity, third-party consensus, and the risk of being wrong about a commercial claim.
That asymmetry has a practical consequence. Content written purely to be cited looks like a reference page: definitions, comparisons, decision criteria, limits. Content written to be recommended needs brand signals that live mostly off the page. A GEO programme that produces only reference content will earn citations and few recommendations, which is exactly the pattern we measured.
The signals we can actually change
Some inputs are testable and within reach. Four of them carry most of the weight in our work.
Entity consistency
The company name, founding year, address and contact details should be identical on the website, in the business registry, and in any third-party listing an assistant might read. We were founded in 2017, and that year appears the same way everywhere it appears. When sources disagree, an answer engine either picks one or hedges, and neither helps.
Extractable answers
An answer engine lifts sentences. A paragraph that opens with a direct answer to the question in its heading is easier to lift than a paragraph that warms up for three sentences. Comparison tables, short definitions, and explicit limits all make the page quotable. This is formatting, not manipulation, and it costs nothing.
The source surface the engine prefers
In our Chinese sample the most-cited domain was Zhihu’s column platform, cited more often than Google’s own developer documentation. In English, community platforms are a well-documented heavyweight. The principle transfers between markets; the list of names does not. If you want to be quoted, you need a presence where the engine in your market already looks.
The indexing base
Before any of the above, the page has to be findable. URL form, canonical tags, reciprocal hreflang, sitemap freshness and internal links are the same checks we run in a standard audit, and they gate everything GEO adds. Google publishes its own guidance for AI features and your site, and the short version is that there is no separate submission route — the normal index is the entry point.
Content patterns that get quoted
Across the answers we’ve annotated, the same few page shapes show up as sources again and again. None of them is exotic, and every one is a formatting decision rather than a copywriting trick.
- A direct answer in the first sentence under the heading. If the heading asks “what is GEO”, the next sentence should define it. Answers are assembled from liftable spans, and a span that begins with throat-clearing is harder to lift.
- A table for any comparison. Options, criteria, thresholds, costs. Tables survive extraction well because the relationships are already explicit.
- Named entities and dates. Company names, product names, a publication date, a measurement date. A page with no dates is hard to cite without hedging.
- Units and numbers with their source. “LCP under 2.5 seconds” beats “fast”. A number with a link to where it came from is easier to trust than an adjective.
- An explicit limits section. Saying what the data doesn’t cover makes a page more usable as a source, not less. Our own studies publish their limits for this reason.
The patterns that don’t get quoted are the ones buyers skim past: “we are a leading provider”, “committed to excellence”, “one-stop solution”. Those phrases carry no extractable fact. They also appear on thousands of competing pages, so they give an answer engine no reason to prefer yours.
What we’d check before spending on GEO
GEO work is easy to sell and hard to verify, so the first month should be diagnostic. Four prerequisites decide whether the later months mean anything.
Search Console access comes first. Without it, indexing rate, query ownership and crawl behaviour are guesses, and every GEO claim sits on sand. Second, the indexed share of the pages that matter. If product or service pages are indexed at a much lower rate than blog posts, that gap is the project. Third, entity consistency across the site, the business registry and third-party listings — it’s the cheapest fix on the list. Fourth, a baseline panel run before any content changes, so the first report has something honest to compare against.
What we would not do first is commission a content calendar. Publishing into a site that answer engines can’t index, or can’t match to a real company, produces pages that never get read. Tianwenseo builds and maintains export sites, and the projects that move fastest are the ones where the technical base was fixed before the first new article was written.
What we can’t measure, and won’t promise
Chat assistants are non-deterministic. The same prompt on two days can produce different sources, and a model update can move every number overnight without warning. Nobody outside the model owner controls that.
We also can’t promise that an AI will recommend you. Our own test is the clearest evidence against that promise: three citations, zero recommendations. Any agency that guarantees an assistant will name your company is selling something it cannot observe, let alone deliver.
Third, GEO is not A/B testable the way an ad is. You can’t isolate one content change and attribute a citation shift to it, because the engine changed too. The workable standard is a monthly panel, a fixed rubric, and a direction of travel over several months.
A GEO scorecard you can run monthly
| Metric | How it’s measured | What a change means |
|---|---|---|
| AI answer coverage | Share of panel prompts returning an AI answer | Market-level shift, not your performance |
| Citation rate | Share of answers listing your domain | Your pages are being read and used |
| Recommendation rate | Share of answers naming you as an option | Brand and consensus signals, slower to move |
| Cited page mix | Which URLs get cited, by cluster | Where to publish next |
| Entity match | Name, year, address, phone identical across sources | Cheap fix with immediate effect |
Five numbers, one page, re-run on the same day each month. That is enough to tell whether the work is doing anything. It will not produce a single score, and that’s the point.
Start from the index, not from the prompt
GEO inherits every technical fault an SEO audit finds, and it adds a measurement layer on top. If pages aren’t indexed, if slugs render as percent-encoded strings, or if the canonical points at the wrong language, the answer engines are working from a partial copy of the site. Fix that first — the indexing test on our own archive shows how misleading a healthy-looking crawl can be.
Then run the panel, keep the raw answers, and report coverage, citation and recommendation separately. The studies behind this page are indexed on the research index, and the technical checks we run before any GEO work are in the 12-point audit.
Questions about generative engine optimization
Is GEO different from SEO?
It adds a measurement layer on top rather than replacing the base. Answer engines still need to crawl, index and render the page, so canonical tags, URL form, hreflang and internal links gate everything. GEO then asks whether an answer appears, whether it cites you, and whether it recommends you.
Can you guarantee that ChatGPT will recommend my company?
No, and nobody can. In our own 72-question test, AI answers named 18 companies and cited our sites three times, each time as a how-to-choose list source rather than as a recommended provider. Citations and recommendations are separate outcomes, and the second one is not under our control.
What tools measure AI visibility?
Panel-based measurement, not a single dashboard. We write 50 to 100 real buyer questions per market, run them logged out in one region and one interface language on the same day, and record four fields: whether an answer appeared, which domains it cited, which companies it named, and in what role.
How often does AI Overview coverage change?
It moves, so a single run is a point measurement rather than a trend. Re-run the same panel monthly and compare like with like. In our Chinese-language sample of 500 queries, 80.9% returned an AI Overview, with coverage higher on conceptual terms and lower on technical lookups.
Does schema markup make AI cite my page?
It helps an engine understand what the page is about, but it does not create a citation on its own. The page still needs to be indexed, to answer a question directly, and to exist somewhere the engine already looks for sources in that market.
Who wrote this and how to check it. Written by Liu Yu, founder of Tianwenseo (Henan Tianwenseo Technology Co., Ltd., registered 3 May 2017; Zhengzhou, China; UTC+8). Tianwenseo has built and maintained WordPress export websites and run Google SEO for B2B manufacturers since 2017, and currently maintains dozens of client sites plus its own research domains. Every figure on this page is measured on a site we run, or linked to the source it came from. We don’t promise rankings or a number of enquiries. Contact: WhatsApp +86 135 2681 6415.