AI SEO Analysis: How to Analyze AI Visibility Engine by Engine

An AI SEO analysis asks one question per engine: when a buyer asks about your category, does this engine name you, describe you correctly, and point to your pages?

The answer depends on four stages, and each one fails differently:

• Parametric. What the model knows with no web access. It is frozen at a training cutoff. Nobody can change it this quarter.
• Access. Whether the engine's crawler may fetch and store your page. You and your CDN control this.
• Retrieval. Whether your page enters the model's context for a given query. The engine controls this, partly.
• Generation. Which retrieved pages the model uses and cites. The model controls this.

Most tools report one blended score across all four stages and every engine. That number cannot tell you what to fix. A useful analysis keeps the stages and the engines apart, so a low result points to a specific cause. If you want to know why two tools give different numbers for the same brand, this is usually why: see why AI visibility tools disagree.

Retrieval is per query and discretionary on every major platform, as of October 2026. ChatGPT decides for itself whether to search. Claude.ai search is a user toggle with no published default. Perplexity's Agent API requires web search to be switched on. Google AI Overviews often do not trigger at all.

That matters because the two kinds of answer come from different places:

• Grounded answers use pages fetched for that query. Changes to your site can reach them once they are crawled and indexed.
• Ungrounded answers come from model memory. OpenAI lists a February 16, 2026 knowledge cutoff for its GPT-5.6 models and August 31, 2025 for the Instant path behind chat-latest. Nothing you publish after the cutoff reaches these answers.

Published estimates of how often ChatGPT searches disagree widely: Profound reports about 18% of conversations, while cloro reports 80.5% on its own prompt set. Those measure different things, so do not assume a default.

How to label answers

1. If the answer shows citations or source links, mark it grounded.
2. If it shows none, mark it ungrounded or unknown. Do not guess.
3. Report the two groups separately. A wrong price in an ungrounded answer is a cutoff problem. The same error in a grounded answer points to a page the engine read.

Also check which model you are testing. Consumer apps route paid and free users to different models, so read whether tools test the same models people use in ChatGPT before trusting a result.

Each engine draws on a different index and exposes different evidence. Google states that AI Overviews and AI Mode "may use different models and techniques" and can return different links, so even two Google surfaces are not one column.

Per-engine results can diverge sharply. In Otterly's 2026 experiment, 2,718 of 3,934 tracked citations for one test site came from Microsoft Copilot, while ChatGPT, Gemini, AI Overviews and AI Mode almost never cited it. A single averaged score would have hidden that.

The same test shows the limit of prompt sampling: Otterly reports that Microsoft's own AI Performance report counted 42,817 citations for that site. Where an engine publishes its own data, use it. For Copilot, start by exporting Bing grounding queries. For Google, see Google's own tools for AI search visibility.

A citation is the link the engine chose to show. It is not a record of everything the model read or used.

Retrieved sources outnumber cited ones

OpenAI's API documentation says the full list of URLs the model consulted is returned separately from citations, and that "the number of sources is often greater than the number of citations." Anthropic and Perplexity also expose retrieval results alongside citations. So a page can be read, shape the answer, and still go uncredited.

Citation behavior is partly separate from source use

Mechanistic research on how language models cite (arXiv 2606.28358) found that amplifying specific internal attention components repaired over 90% of missed citations without changing the sources. In other words, whether a citation appears depends partly on internal model behavior, not only on which pages influenced the answer.

Citation units differ by engine

Anthropic cites short text spans. Perplexity cites whole sources by number. Adding those together produces a number with no shared denominator.

What this means for your analysis:

• Read citation counts as a floor, not a total.
• Where an API exposes retrieved sources, log them next to citations. The gap between retrieved and cited is often more informative than either count.
• Check the accuracy of what the answer says about you, not only whether you were linked. A wrong fact without a citation still reaches the buyer.

The levers with solid evidence behind them are access levers. They work by removing blocks, not by adding signals.

In your control:

• Per-bot robots.txt rules. OpenAI's crawler page says each setting is independent. OAI-SearchBot governs ChatGPT search, while GPTBot governs training only.
• Network reachability. A firewall or CDN can block a crawler that robots.txt allows. Reading server logs for AI crawler status codes shows whether bots get a 200 or a 403.
• Server-rendered content. Most AI crawlers do not execute JavaScript, so facts that appear only after hydration may be invisible to them. Googlebot is the main exception.
• Index and snippet eligibility. Google requires a page to be indexed and eligible for a snippet to appear as a supporting link in AI Overviews or AI Mode. noindex, nosnippet, max-snippet:0 and data-nosnippet keep content out of AI answers.
• Clear, fetchable facts. Stating a price or a one-line definition on a crawlable page gives an engine a sentence it can quote.

Not in your control:

• What the model learned before its cutoff.
• Whether an engine searches for a given query.
• Your position in the model's context list.
• Which retrieved pages it chooses to cite.

A July 2026 critical survey of 45 GEO studies (arXiv 2607.14035) found that no reviewed technique shows a stable, cross-platform causal effect on discoverability. Be wary of any analysis that promises a score or a ranking.

Run the analysis in the order the stages occur. A later stage cannot succeed if an earlier one failed.

1. Write a short set of buyer questions. Use the questions real prospects ask. Remember that engines rewrite prompts, so these are a proxy: see whether tools can see the prompts people actually type.
2. Take a baseline snapshot. The free checker on this site shows whether engines name you for up to three buyer questions. Treat it as a snapshot, not tracking.
3. Check access per crawler. Compare robots.txt with what each crawler actually receives. The AI Crawler Checker requests a page as each AI crawler and shows which ones the firewall turns away.
4. Diff raw HTML against the rendered page. List any price, product name or claim that appears only after JavaScript runs.
5. Check index status engine by engine. Look at Google Search Console and Bing Webmaster Tools separately. One being healthy tells you nothing about the other.
6. Run each question per engine, logged out, with memory off. Label each answer grounded or ungrounded. Record citations, and retrieved sources where an API exposes them.
7. Score accuracy. Compare what each answer states with your verified facts. Flag wrong prices, outdated product names and competitor confusion.
8. Benchmark against like-sized brands. Brand size dominates AI citation patterns, so benchmark against like-sized competitors rather than category giants.
9. Fix the earliest broken stage first, then re-run. Allow time for crawling before you compare.

• Blending engines into one score. It averages answers with and without retrieval across engines that share no denominator. Report per engine.
• Reading citation as influence. Retrieved sources outnumber cited ones, so an uncited page may still have shaped the answer.
• Tracking an "AI rank" in Google. Google's Search Console report for AI features shows impressions only. No position data exists to check such a number against.
• Testing logged in. OpenAI says ChatGPT may use saved memories when rewriting a search query. Personal context makes results hard to reproduce.
• Treating your prompt list as real demand. Engines generate their own retrieval queries. Bing grounding queries are the closest public view of the real input.
• Comparing scores across method changes. If your tool changes its formula or models, old and new scores are not comparable.
• Chasing tactics the evidence does not support. Schema, llms.txt, E-E-A-T scores, chunking and IndexNow pings are not proven routes into AI answers. Ahrefs' matched test found AI Overview citations fell 4.6% after pages added schema.
• Trying to manipulate answers. Hidden text, prompt injection and cloaking for AI crawlers are illegitimate. Research on search agents shows manipulated content can sway answers, which is a risk to your brand when others do it, not a tactic to copy.

What is the difference between a tracking tool and an AI SEO analysis? A tracking tool reports how often engines mention or cite you over time. An analysis explains why, by checking each stage: crawler access, rendering, index status, retrieval and the accuracy of what the answer says. Tracking tells you a number moved. Analysis tells you which block to remove. Many teams need both, but only the analysis points to a fix you can ship.

How can I tell if an AI answer used live retrieval? The clearest sign is visible citations or source links in the answer. If none appear, you usually cannot tell whether the engine searched, so label it unknown rather than guessing. On APIs you can control retrieval directly by switching web search on or off, which lets you compare what an engine says about your brand from memory against what it says after reading the web.

Why does my brand appear in Copilot but not ChatGPT? Each engine draws on a different index and makes its own retrieval and citation choices. Copilot grounds on Bing, so Bing index status matters most there. In Otterly's 2026 experiment, ranking in Bing produced effectively zero ChatGPT citations even while ChatGPT's crawlers visited the sites. Check access and index status per engine rather than assuming one feeds another.

How often should I re-run an AI SEO analysis? Re-run after you ship a fix that affects access, rendering or indexing, once crawlers have had time to revisit. Also re-run when an engine changes its default model, because a new model can change what ungrounded answers say. Between those events, a regular cadence catches drift. Keep the same prompts and settings each time so results stay comparable.

Can an AI SEO analysis guarantee more citations? No. An analysis can show what blocks you, such as a firewall returning 403 to a crawler or facts that only load with JavaScript, and fixing those makes you eligible. It cannot control whether an engine searches, which pages it retrieves, or which it cites. Any tool that promises a citation count or a ranking is promising something the engines do not let anyone control.