Which GEO ranking factors actually work?

Only three GEO ranking factors hold up in large studies. Rank in Google and Bing for the sub-questions AI tools search for. Give each page a title that matches the question. And get your brand named on other sites. Schema markup, llms.txt files, E-E-A-T and readability scores show no measured lift in AI citations so far.

How did I check each factor?

Most GEO factor lists are SEO advice with a new label. Few of them say where a factor comes from. So I went back to the research behind each one and sorted it by the kind of evidence.

Controlled tests change one thing and compare against pages that didn't change. They're the strongest kind, and the rarest.

Large correlations look for patterns in millions of citations. They show direction, and they can't prove cause.

Lab benchmarks rewrite pages and test them against a model. They're useful, and a long way from live search.

Google's own guide says the usual SEO practices still apply to AI Overviews and AI Mode. It treats GEO and AEO as SEO for AI search¹. One more thing to know up front. Most of the large datasets come from Ahrefs, which sells AI visibility tracking. The studies are careful, and I note it anyway.

Which GEO factors have evidence behind them?

Does ranking in Google and Bing still matter?

Yes, it's still the way in. What changed is which queries count.

In July 2025, Ahrefs found that 76% of pages cited in AI Overviews ranked in Google's top 10 for the same query². By March 2026, across 4 million cited URLs, that share was 38%. Another 31% ranked 11 to 100, and 31% didn't rank in the top 100 at all³.

That looks like rankings stopped mattering. They didn't. Ahrefs puts most of the gap down to query fan-out³. Google turns one question into several related searches and cites pages that rank for those¹. Ahrefs also improved its parsing between the two studies, so the numbers aren't a clean comparison³.

ChatGPT works the same way. In Ahrefs' study of 1.4 million prompts, 88% of the pages ChatGPT cited came through its web search channel⁴. Seer Interactive matched more than 87% of ChatGPT's citations to Bing's top 20 results, in a small test of about 100 questions⁵.

The toughest independent test agrees. In C-SEO Bench, a NeurIPS 2025 benchmark, methods that moved a page higher in the model's list of sources beat AI-specific rewriting by a clear margin⁶.

Why does ChatGPT cite one page and skip another?

Mostly because of the title. ChatGPT retrieves about twice as many pages as it cites. In Ahrefs' data, 49.98% of retrieved pages got a citation⁴.

The source matters first. Pages that came from web search were cited 88.46% of the time. Reddit pages were cited 1.93% of the time, yet they made up 67.8% of everything ChatGPT read and dropped⁴. It reads the crowd to learn the consensus, then credits someone else.

Within search results, the title decides. Cited titles scored 0.602 in similarity to the prompt, against 0.484 for skipped ones. Against ChatGPT's own sub-queries, the best match rose to 0.656⁴. The URL helps too. Readable slugs were cited 89.78% of the time, against 81.11% for coded ones⁴.

So the "topical depth" in most factor lists really means this. Learn the sub-questions behind a buyer's question. Then answer each one in a page or section whose title says it in plain words. Google warns against mass-producing a page for every variation, which breaks its scaled content abuse policy¹.

Do brand mentions matter more than backlinks?

In the data, by a wide margin. Ahrefs compared 75,000 brands with how often Google's AI Overviews name them. Branded web mentions correlated at 0.664. Backlinks correlated at 0.218⁷.

The top three signals all live off your site. They're mentions of your brand, your brand name in link text, and people searching for your name⁷. The spread is steep. Brands in the top quarter for web mentions averaged 169 AI Overview mentions. The next quarter averaged 14⁷.

Keep two warnings in mind. This is correlation, and Ahrefs says so. Famous brands get mentioned and cited for the same reasons. And Google lists chasing inauthentic mentions as a tactic to skip¹. Earn mentions with work people want to talk about.

Does schema markup get you cited?

It hasn't in the only large controlled test. AI-cited pages were almost three times as likely to carry JSON-LD as uncited pages. That's the number in most GEO decks⁸.

So Ahrefs tested it. They tracked 1,885 pages that added JSON-LD between August 2025 and March 2026. Each was matched with three similar pages that didn't add it⁸.

ChatGPT moved by +2.2% and AI Mode by +2.4%, both within noise. AI Overviews fell 4.6%. That drop was significant but small, and both groups were already falling⁸. A searchVIU test cited in the study explains why. ChatGPT, Claude, Perplexity, Gemini and AI Mode read only the visible text when they fetched a page. They skipped JSON-LD entirely⁸.

Google says the same in its guide. Structured data isn't required for its AI features, and it still earns rich results in normal search¹.

The test covered pages that were already well cited, so it can't speak for new pages. I still add Organization and Person schema to every page on this site. I do it so the facts about who I am match everywhere, which is the job of a context system. I don't expect it to win citations by itself.

Does fresh content get cited more?

A little, and mostly as a tie-breaker. Across 17 million citations, pages cited by AI assistants averaged 1,064 days old. Google's organic results averaged 1,432 days. That makes AI citations 25.7% fresher⁹.

ChatGPT leans furthest toward new pages. Google's AI Overviews don't lean at all, and cite pages about 16 days older than organic results⁹. In the ChatGPT study, age mattered mainly when two pages were equally relevant⁴. The average cited page is still almost three years old⁹.

Update a page when its facts change, and change the date only then. Google's John Mueller has warned against new dates on unchanged pages⁹.

Is E-E-A-T a ranking factor?

No, and Google says so in writing. Its guidance states that E-E-A-T itself isn't a specific ranking factor. Google uses a mix of other signals to find content that shows it¹⁰.

Google's AI guide doesn't use the term at all. It asks for first-hand experience and a point of view of your own¹. So treat E-E-A-T as a description of good content. Real authors, real sources and real experience help readers trust you, and that's reason enough to show them.

Do statistics and quotes raise AI visibility?

In the lab, yes. In a tougher benchmark, mostly no. The original GEO paper tested nine rewrites on 10,000 queries. Adding quotations, statistics and cited sources raised visibility by up to 40%, and by up to 37% on Perplexity¹¹. Keyword stuffing barely moved.

C-SEO Bench retested methods like these across six domains, with several sites competing at once. Most were largely ineffective, and often pushed pages down. The gains also shrank as more sites used them⁶.

So add numbers because your readers need them. A page with original data is the page other sites mention, and mentions are the factor that holds up.

Does an llms.txt file help?

Almost nothing reads it. Ahrefs checked the server logs of 137,210 sites in May 2026. 28% published an llms.txt file. Of those 38,360 files, 97% got zero requests all month¹².

Google's guide says Google Search doesn't use llms.txt or other special AI files¹. This site has one anyway. It took ten minutes and costs nothing. I don't count it as a ranking factor.

What about readability scores and engagement?

I found no study that links them to AI citations. AI tools never see your bounce rate or your Flesch score. They see your text.

Clear structure still helps in one proven way. Models use information at the start and end of their context best, and lose track of the middle¹³. So put the answer first, in plain words, and let the rest of the page back it up. Google adds that there's no need to cut content into tiny chunks for AI¹.

What should you work on first?

Here's the order I'd follow, strongest evidence first.

Map the sub-questions. Write down 20 to 50 questions buyers ask on the way to your service. Check which of your pages rank for them in Google and Bing. My guide to getting cited by ChatGPT covers the page changes.

Title every page as an answer. The title and URL should say the question or the answer, in your buyers' words.

Earn mentions on other sites. Reviews, podcasts, YouTube, industry lists and data other people can quote all count.

Publish what nobody else has. Your own numbers, real examples and named sources give others a reason to mention you.

Keep the answer in visible text. AI tools read the text on the page. Facts that live only in markup or scripts get skipped⁸.

Track share of answers. Run the same buyer prompts every month and record who gets named. That's AI answer monitoring, and it's why AI citations behave like impressions.

To plan those pages, start with the free AI website brief builder.

Questions people ask