What AI SEO Research Actually Shows

Four pieces of work carry most of the weight in GEO research. Each found something real. Each measured something narrower than it is usually quoted for.

The GEO paper

The original GEO paper is where the "up to 40%" figure comes from. It reports a 30 to 40% band across three methods, measured with Position-Adjusted Word Count. Its one "in the wild" Perplexity test uploaded source text as files, because Perplexity would not accept source URLs. The same paper found keyword stuffing lowered its metric (17.8 against a 19.5 baseline).

C-SEO Bench

C-SEO Bench concluded that "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking." It also found gains shrink as more sites adopt the same tactics, which it called "congested and zero-sum."

The factorial and the survey

The 252,000-trial factorial found that "formatting-only edits have little impact." The critical survey of 45 GEO studies reached the governing conclusion: already-retrieved content can change its citation, but no reviewed technique moves organic discoverability in a stable way.

An AI answer passes through four stages. Mixing them up is the main source of over-claiming in GEO research.

1. Parametric: what the model knows with no web access, frozen at a training cutoff. Nobody changes this quickly.
2. Access: whether a crawler may fetch and store your page. You and your CDN control this.
3. Retrieval: whether your page enters the model's context for a given query. The engine controls this, partly through its index.
4. Generation: which retrieved documents the model uses and cites. The model controls this.

The GEO paper, C-SEO Bench and the factorial all start at stage 4. The document is already in the context window. That skips the two gates most brands actually fail: being fetchable, and being retrieved.

A rewrite that lifts citation among two pre-loaded candidates tells you nothing about whether a live engine would ever retrieve the page. As of October 2026, retrieval is also discretionary. OpenAI says ChatGPT search rewrites a prompt into one or more targeted queries, so the query that decides retrieval isn't the one a study typed. For a fuller comparison, see lab GEO studies vs live engines.

The findings above survive scrutiny better than most GEO claims. They still carry limits a reader should keep in view.

• Position is not a lever. C-SEO Bench's "traditional SEO" arm was implemented by moving the target document to the top of the context. No index was queried and no ranking was earned.
• Small candidate pools. The factorial used two documents per trial, which likely overstates position effects compared with production pools.
• Interested authors. The factorial is an industry submission that ships a GEO checklist. The survey is single-author and unaffiliated. Two interested sources agreeing is corroboration, not replication.
• Dated models. The rewrite benchmarks are 2025 work with no replication on the current frontier models.
• A within-context null is not a total null. These studies don't prove content quality is irrelevant. They don't test what helps a page enter the consideration set.

The schema evidence has the same boundary. Ahrefs tracked 1,885 pages adding schema against matched controls and found AI Overviews citations fell 4.6%, but every treated page already had 100+ AI Overview citations. Whether schema helps a page get considered at all is untested, not refuted. Different setups like these explain most of why GEO studies disagree.

The research points one way: the levers with a documented mechanism remove blocks. None of them add a boost.

In your control

• Rendering. Vercel found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so content that appears only after JavaScript runs may be invisible to them.
• Per-bot robots.txt. OpenAI's GPTBot governs training. OAI-SearchBot governs ChatGPT search. Each setting is independent.
• Network reachability. A firewall or CDN can block an AI crawler that robots.txt allows.
• Index and snippet eligibility. noindex, nosnippet, max-snippet:0 and data-nosnippet keep content out of AI answers.
• Specific facts on fetchable pages. A stated price, number or one-line definition gives an engine a sentence it can quote. The factorial found explicit price information helped consistently.

Not in your control

• Your position in the engine's candidate list.
• The queries the engine actually runs.
• What the model learned before its cutoff.
• How prominent your brand already is, which is associated with citation in every correlational study.

Being named on the sites engines already cite is associated with appearing in answers. Nobody has measured a guaranteed effect. The evidence for each side is laid out in on-page rewrites vs access fixes.

Start where the evidence is strongest and the check is cheapest.

1. Check crawler access. Confirm robots.txt allows the retrieval bots you care about: OAI-SearchBot, PerplexityBot, Claude-SearchBot. Then confirm your firewall returns a 200 to them. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
2. Diff raw HTML against the rendered page. Find the prices, product names and claims that only appear after JavaScript runs. Server-render those.
3. Audit suppression controls. Look for template-wide noindex or data-nosnippet that hides content by accident.
4. Check index status per engine. Use Google Search Console and Bing Webmaster Tools separately. One being healthy says nothing about the other.
5. State key facts plainly and consistently. Put your pricing and core claims in fetchable text, and make every page agree.
6. See what engines say. The free checker on this site shows whether AI platforms name you for three buyer questions. Treat it as a snapshot, not a trend.

A free GEO audit follows the same order: access first, then accuracy.

• Quoting the 40% figure as a live result. It came from a lab engine with sources pre-loaded into context.
• Reading position effects as proof that SEO drives AI citations. The studies simulated top position. They didn't earn it.
• Treating citation as use. Engines retrieve more sources than they cite, and research on attribution shows citation can be changed without changing the sources.
• Selling schema, llms.txt or chunking as citation levers. Google says Search ignores llms.txt, structured data isn't required for generative AI search, and there is no ideal page length.
• Trusting a before-and-after with no control group. Brand size confounds most correlational findings. Read correlation vs causation in AI search studies before acting on one.

The fix for all five is the same question: which stage did this study actually measure?

Does this research mean GEO is useless? No. It means the published evidence supports a narrower claim than the market sells. Rewrites can change how an already-retrieved page gets cited. They have no demonstrated effect on whether a live engine finds you. The work that holds up is access: rendering, crawler permissions, firewall reachability and index eligibility. Those are mechanical, checkable and in your control.

Do GEO rewrites ever hurt? They can. C-SEO Bench found most conversational SEO methods frequently lowered document ranking, the opposite of what was expected. The 2026 survey also warns that citation-oriented rewrites can impair retrieval. Edits that inject statistics, boilerplate authority phrases or keyword density are better treated as risks than wins.

Is there any positive on-page signal in the research? One. The 252,000-trial factorial found that including explicit price information and a recent timestamp helped consistently. That fits a simple mechanism: a specific fact on a fetchable page gives the engine a sentence it can quote. It was still measured inside the model's context, so it says nothing about retrieval.

How should I judge a new GEO study? Ask four things. Did it query a live engine, or supply documents in context? Did it use a control group? Who funded it, and do they sell the tactic? Does the metric count something the intervention itself added? A study that fails these can still be interesting. It just can't tell you what will happen to your brand in a real engine.