How to Evaluate AEO Research Before Changing Your SEO

Ahrefs checked more than 17,000 pages to find who publishes the most original AI search research. The result, a list of 25 people, is a handy map of who to read. It's a poor shortcut for deciding what to believe.

Studies keep arriving with different sample sizes, methods and motives, and some seem to contradict each other. This guide gives you a repeatable way to grade any of them, walks through an audit of one real claim, and shows how to test a finding on your own site before you change anything.

What the list tells you, and what it doesn't

The author, Ryan Law, started with 87 names: people widely recommended as AEO and GEO experts, plus researchers at AI visibility companies. He counted original AI search studies per person since 2023, ranked by that count, and broke ties by the number of sites linking to those studies. Company research with no named author counted toward nobody.

So the ranking measures volume. It doesn't measure accuracy, rigor or independence. A person with 36 studies has published more than a person with 4, and that's all the table proves.

Look at the company column too. Many of the listed authors work for software or agency businesses: Ahrefs, Peec AI, Profound, Search Atlas, SISTRIX, seoClarity and others. Law works at Ahrefs, and Ahrefs staff hold ranks 2, 4 and 9. That isn't a sign of bad faith. It's a reason to read the list as a directory of publishers, not a ranking of truth.

Step 1: Run the 7-point checklist

The list used four tests to decide what counted as a study: original data, an AI-search focus, a meaningful sample and a published write-up. That's a sensible entry filter, but it says little about quality. The checklist below keeps the first three, drops the fourth (a study you can read is a given), and adds four checks the list doesn't cover.

Score each point pass, caution or fail.

1. Original data. Pass: the authors collected or analyzed the data themselves (prompt runs, citation datasets, server logs, analytics, surveys, experiments). Caution: the study re-slices someone else's numbers or doesn't say where the data came from. Fail: opinion or a recycled statistic with no dataset behind it.
2. AI-search focus. Pass: the data measures the exact behavior the headline claims, on named platforms. Caution: the data covers one surface, such as AI Overviews, but the headline says "AI search" in general. Fail: the conclusion is inferred from traditional rankings alone.
3. Sample and setup. The list's floor is roughly 50+ prompts, queries or pages, analytics from 10+ sites, 100+ survey respondents, or a controlled experiment with a described setup. Treat that as the minimum, not a mark of strength. Pass: well above the floor, with prompt selection, platforms and dates stated. Caution: near the floor or vague on selection. Fail: below the floor or undescribed.
4. Funding and incentive. Pass: the sponsor is disclosed and the finding doesn't help sell anything. Caution: the publisher is a vendor and the finding fits its product. Fail: an undisclosed sponsor plus a conclusion that directly sells the sponsor's offer. A vendor isn't automatically wrong, but you should know who benefits.
5. Replicability. Pass: the prompt list, page set or method is published, or detailed enough to rerun. Caution: the method is described only in loose prose. Fail: "proprietary data" with no method.
6. Age of the data. Pass: collected recently and dated. Caution: several months old. That's a rule of thumb, but age matters here because the list points to researchers measuring heavy change in cited sources over time (Profound's volatility study, SISTRIX's citation drift study). Fail: undated.
7. Agreement with other studies. Pass: at least one independent dataset points the same way. Caution: only follow-ups from the same team agree. Fail: a stronger study contradicts it and nobody explains the gap.

Turn the scores into a decision

Any fail on points 1 to 3 means ignore the study as evidence. Otherwise, count your cautions. This pairs well with see also: how to audit content for chatgpt information seeking.

• Zero or one caution: a candidate to adopt, ideally after a light check if the change is costly.
• Two to four cautions: test it on your own site first.
• Five or more: park it and wait for better data.

These thresholds are a judgment call, not a standard. Adjust them to how hard the change would be to undo.

Step 2: Audit a real claim

Take a claim that touches a lot of technical SEO budgets: schema markup and llms.txt do little for AI citations. This audit uses only what the Ahrefs list states about it. That page summarizes the studies rather than reproducing them.

What the page says. Louise Linehan's Ahrefs work includes a test in which 1,885 pages that added schema saw barely any change in AI citations, and a finding that 97% of llms.txt files never get read (137,210 domains). Manick Bhan's Search Atlas research, built on datasets such as 120M+ AI citations across 6.4M domains, reports no measurable effect from schema markup or /llms.txt on visibility. Dan Petrovic's DEJAN published a log observation that Google, Anthropic and OpenAI crawlers fetched dejan.ai's robots.txt 1,150 times and its llms.txt 4 times in 30 days.

Scoring what the page lets you score.

1. Original data: pass. Each team describes its own datasets, tests or logs.
2. AI-search focus: pass. All three concern AI citations, visibility or AI crawler behavior. The DEJAN observation, though, measures crawler fetches on one site, not citations.
3. Sample and setup: pass on size, caution on setup. 1,885 pages sits far above the 50-page floor, and the citation datasets are huge. But the Ahrefs list doesn't say how the schema test was set up: whether it had a control group, how long it ran, or which platforms it checked.
4. Funding and incentive: caution. Linehan works at Ahrefs, Bhan runs Search Atlas, and Petrovic runs DEJAN. All three publish as companies. Whether a null result on schema helps or hurts their business is unclear, so that's an inference, not a finding.
5. Replicability: unknown. The list doesn't say whether prompts and page lists are published. You need the full studies.
6. Age: unknown. The list gives no collection dates for these studies, only that it covers January 2023 to November 2026.
7. Agreement: pass, with a caveat. Two separate teams on different datasets report little or no effect, and the DEJAN logs fit the llms.txt story. Still, all three sources are companies, and they measure different things (reads, fetches, citations).

Verdict. Two cautions and two unknowns, no fails. By the rule above, that lands in "test on your own site." It's reasonable to hold off on making schema or llms.txt a citation-driven project. Nothing here justifies removing existing markup, because these studies measure AI citations only and say nothing about other uses of schema. This pairs well with ai performance review questions for seo teams that work in depth.

Limits of this audit. It rests on summaries, not the full studies, so points 5 and 6 stay open until you read the originals. And a "no measurable effect" result can't prove there's no effect. It only shows that none appeared at that scale and setup.

Step 3: Reconcile studies that seem to conflict

Before you pick a side, line up four things: the surface (AI Overviews, ChatGPT, Perplexity), the prompt or query set, the collection date, and the exact metric.

The list itself offers an example. Ahrefs highlights "Update: 38% of AI Overview Citations Pull From The Top 10" and "Only 12% of AI Cited URLs Rank in Google's Top 10 for the Original Prompt." The numbers look far apart, but the list gives only the titles, not the full definitions. A gap that size could come from a different surface, a different match rule or a different prompt set. That's an inference, so check those details in the studies before you call it a conflict.

Also read: also worth reading: noindex vs robots.txt for shared pages: how to audit

If the definitions match and the numbers still disagree, trust the study with the larger sample, the described setup and the more recent data. If that's unclear, treat the question as open and test it yourself.

Step 4: Verify the claim on your own pages

A small test can't match a 1,885-page study, so it will only catch large effects. It still beats acting on faith.

1. Pick a fixed set. Choose 50+ prompts your buyers would plausibly ask, or a set of 50+ pages. For a page-level change, split the pages into a test group and an untouched control group.
2. Record a baseline. For each prompt, note whether your domain is cited, which URL, and on which platform. Run each prompt more than once, since the list points to research showing cited sources change over time.
3. Change one variable. Add schema to the test pages only, for example. Don't touch titles, copy or links on those pages during the test.
4. Set the window in advance. Pick a re-check date, such as four weeks out, and write it down. The length is your call, not a standard.
5. Re-run and compare. Repeat the same prompts the same way, then compare the test group's change with the control group's. A gap that matches the noise between your baseline runs isn't a result.
6. Log what you can't rule out. Seasonality, new content, model updates and competitor moves all muddy the result. Write them next to your numbers.

A flat result matches what the published studies report, and that's useful: it tells you not to spend more on that variable for citations. A large, repeatable gap is worth rerunning on a bigger set before you roll it out.

Keep a running sheet

Sample size and setup are easy to read in most studies. Funding, age and replicability usually take a trip to the full write-up. Keep a simple sheet with one row per study, the seven scores and your decision, and reuse it whenever a new report lands in your feed.