Generative engine optimization, measured honestly.

What generative engine optimization is.

Generative engine optimization, or GEO, is the practice of making a brand more likely to be named and cited when an AI model answers a question in its category. A generative engine is any system that writes an answer instead of listing results: the AI assistants, and the AI summaries that now sit above ordinary search results.

The term comes from a 2023 research paper, GEO: Generative Engine Optimization by Aggarwal and colleagues, later accepted to KDD 2024. The authors built a benchmark of queries and tested ways of rewriting source pages. Citing sources, adding quotations and adding statistics worked best, raising a page’s visibility in generated answers by up to 40 percent on their own measure. Keyword stuffing did not help. They also found that what worked varied by domain. That is a useful starting point, and a reminder that it was measured mostly in a research setup rather than on the assistants your buyers use today.

You will see other names for the same idea. AEO, answer engine optimization, predates large language models and began with featured snippets and voice search. LLMO, large language model optimization, and AI SEO are also used. The differences are mostly emphasis. The short definitions are in the glossary under GEO, AEO and LLMO.

How it differs from SEO.

SEO has a clear result to work toward: a position for a query, which you can look up and which stays put long enough to report. GEO has no single equivalent, and four differences explain why.

• The output is prose. An answer can name you, link to you, do both, or do neither, and each of those is a different result.
• The answer varies. Ask the same model the same question twice and the brands it names can change. One answer is a sample, not a ranking.
• There are two sources of knowledge. A model answers partly from what it learned in training and, when web search is on, partly from pages it retrieves for that question. The two can disagree about your brand.
• The model matters. Different providers, and different versions from one provider, give different answers to the same question. A number without a model version attached cannot be compared next month.

What carries over is most of the groundwork. Pages that crawlers can reach, that answer a buyer question plainly, and that other sites reference are useful for both. GEO does not replace SEO. It adds a second result to measure, and a different way of measuring it.

How an AI answer chooses its sources.

The details differ by provider and are not fully published, so treat this as the general shape rather than a specification.

Without web search, the model answers from memory.

The model writes from what it learned in training. It has no list of sources to consult, and any link it writes is reconstructed from memory, which is why such links are sometimes dead. Your influence here is slow and indirect: what the public web said about you before the training data was collected.

With web search, it retrieves pages first.

The assistant turns the question into one or more searches, fetches a set of pages, picks the passages that answer the question, and writes an answer from them, usually with links to the pages it used. To be chosen, a page first has to be reachable by that provider’s crawler and present in the index it searches. After that, a page that states the answer clearly in one section is easier to quote than one that buries it.

Third party pages often matter, so check which ones.

For questions such as “which tool should I use for this”, an answer can lean on comparison articles, review sites, forums and directories as much as on the vendors’ own pages. How much varies by category and by model, so check rather than assume: list the pages the models cite for your questions, and you know where it is worth being present.

A citation is not a mention.

These are two separate events, and treating them as one is a common source of confusion in GEO reporting.

• A citation is a page the answer linked to as a source: where the answer says it got its information. See citation share.
• A mention is your brand’s name in the prose of the answer: the text your buyer actually reads. See brand mention.

They come apart in both directions. A model can name you without linking to anything of yours. And it can link to your own page as a source while never saying your name, recommending three competitors in the text instead.

That second case deserves its own place in your reporting. The model found your page, read it, and built part of its answer on it, so being found was not the problem. What is missing is your name in the text. That points at a specific page and a specific piece of work. It is an observation, not a proven cause: nobody can promise that putting your name on that page changes the next answer. It is still the most concrete lead a measurement can give you.

An example. You sell scheduling software, and a model answering “how do small clinics cut no shows” cites your guide on reminder timing, then recommends two other products by name. Open the guide: if your product is named only in the footer, the page may never have told the model who wrote it or what you sell. Say it in the text, where the advice is.

Two practical rules follow. A link to your site should never be counted as a mention, so strip the URLs from the answer before you look for your name. And when a report shows a number, it should say which of the two it is counting. When two dashboards disagree, that is usually the first thing to ask.

How to measure it honestly.

GEO numbers are easy to produce and easy to misread. Five habits keep them honest.

1. Fix the questions.

Write down the questions your buyers ask, in their words, and keep the set stable. Change the set and you change the measurement, so a rise could be the new questions rather than your work. Mix three kinds: category questions (“best scheduling software for small clinics”), comparison questions (“alternatives to the tool we use now”) and problem questions (“how do we cut no shows”). Keep questions that already contain your brand name apart, since those say little about whether a buyer would find you.

2. Name the model version.

Record the provider, the model and the exact version behind every answer. When a provider ships a new version your numbers can move with no change on your side, and without the version you cannot tell the two apart.

3. Keep with search and without search apart.

An answer from memory and an answer grounded in live search are two different measurements. They can differ a great deal for the same brand and the same question. Report them side by side, never added together.

4. Repeat the runs and show the sample size.

Because answers vary, a percentage is only as good as the number of answers behind it. Named in two of three answers is not 67 percent in any useful sense. Run the same question many times, report the count next to every percentage, and where you can, report a confidence interval: the range the true rate plausibly sits in, given how many answers you sampled. A floor such as five answers only stops a number from being published on almost nothing. A tight range takes far more runs than that.

5. Count mentions and citations separately.

Track how often your name appears in the prose, how often your pages are cited, and how often both happen in the same answer. The gap between them is where the work is.

Our own version of these rules, including which numbers carry an interval and where the method is weak, is on how we measure.

What to change.

If you have a week, start with three things: check that the AI search crawlers can reach your important pages, fix the pages that are cited without naming you, and find the third party pages cited for your category. The changes that matter fall into a few groups. The full list, with how to check each item, is in what makes AI search engines cite a page. In short:

• Let the AI search crawlers in. Several providers run separate crawlers for search and for training, so you can often block training and still allow search. Google handles this differently, so read each provider’s documentation. Check your robots rules and your CDN bot settings: a page the search crawler cannot fetch is unlikely to be cited.
• Answer one question per section, near the top, in plain sentences. A passage that stands on its own is easier to quote.
• Back claims with named sources, quotations and real figures, the changes the GEO paper found most effective, and keep pages current, with a visible date when they change.
• Say your name on the pages that get cited. If a model cites your page and still names someone else, check whether the page itself makes clear who you are and what you sell.
• Be present where the models look. Find the third party pages cited for your category and earn an accurate place on them.

Then measure again, with the same questions on the same model versions, and compare like with like.

How Whaily does this.

Whaily is a tool for measuring the things on this page. It calls the model APIs rather than reading the chat apps your buyers use, which is what makes one day comparable with the next. The whole set of numbers is described on AI search analytics. What it does today:

• It runs the questions you choose once a day on the models you choose, from OpenAI, Anthropic, Google and other providers, through their APIs, and stores the provider, model and exact version with every answer. Daily runs give you a trend over time; bursts, below, give you a sample size. Model coverage.
• It counts a mention only when your name is in the prose, with URLs removed first, and records citations as a separate set of pages. AI visibility and citation sources.
• It lists the answers that cited a page on your domain without naming you, and from five such answers up it shows them as a rate.
• It shows no percentage below five answers. Below that, the number reads N/A. Five is a floor, not a promise of precision.
• A burst runs one question many times across the models you pick and returns the mention rate with a 95 percent confidence interval, the tool for when one daily answer per model is not enough. Bursts.
• On the Pro, Agency and Enterprise plans a question can also run with web search on one or two of its models, and the two results are shown side by side, never summed.
• An action plan, rebuilt every night from your stored answers and sources, ranks what to do first. Action plan.

What it does not do: it does not read the ChatGPT app, Google’s AI summaries or any other logged in chat interface, where a system prompt and memory change the answer, and it does not measure traffic that arrives from AI assistants.

Frequently asked questions.

Is generative engine optimization the same as SEO?

No, though they overlap. SEO works toward a position in a list of links. GEO works toward being named, and being cited as a source, inside an answer a model writes. Crawlable, well organised pages help both, but the result you measure is different.

Are GEO, AEO and LLMO different things?

Mostly they are different names for the same work. AEO, answer engine optimization, is older and started with featured snippets and voice search. LLMO, large language model optimization, and AI SEO are used for the same goal. The method matters more than the label.

How long does it take to see results from GEO?

Nobody can promise a timeline. Answers that use live web search can pick up a changed page once the provider has recrawled it, which can take days or weeks. What a model remembers without search changes only when the provider ships a new version. So check search answers a few weeks after a change, check memory answers after each new model version, and compare like with like.

Can I measure GEO by asking ChatGPT myself?

You can look, but one answer is one sample. The same question asked twice can name different brands. To measure, fix the questions, name the model version, repeat the runs, and report how many answers each percentage rests on.

What does it mean when an AI answer cites my page but does not name my brand?

The model used your page as a source and then named other brands, or none, in the text a reader sees. It is worth tracking separately, because the model already found the page. Whaily calls it cited without being named.

Does Whaily measure what people see in the ChatGPT app?

No. Whaily calls the model APIs, which is what makes results comparable across days and versions. A logged in chat app adds its own system prompt, memory and routing, so its answers differ from person to person, and Whaily does not automate those apps.