Free AI Search Simulator: How AI Engines Pick Sources, Stage by Stage

Every AI answer goes through four stages. A simulator handles each one differently.

• Trained knowledge. With search off, the model answers from what it learned before its cutoff. OpenAI's docs list a Feb 16, 2026 cutoff for the GPT-5.6 family and Aug 31, 2025 for the Instant path. Nothing you publish now reaches that answer.
• Access. A crawler has to be allowed in by robots.txt, let through by your firewall or CDN, and able to read your content. This stage can be tested for real.
• Retrieval. The engine picks pages for its own version of the question. You can't observe this, only approximate it.
• Citation. The model chooses which retrieved pages to name. A citation is not proof the page shaped the answer.

Here is how the access and retrieval stages differ by engine, as of October 2026:

For the full path from query to citation, read how AI search picks sources, or the engine pages on where ChatGPT citations come from and how Google picks AI Overview supporting links.

You can run the testable part yourself in about 20 minutes. Replace the domain and page with your own.

1. Read your robots.txt per bot. Find the rules for OAI-SearchBot, PerplexityBot, Claude-SearchBot and Googlebot separately. The common error is a blanket AI block that also removes ChatGPT search eligibility.
2. Request a key page as each crawler. Use the full user-agent string from each vendor's crawler page in place of the short token below. A 200 is a pass. A 403, a 429 or a challenge page is a block.
3. Check the raw HTML for your key facts. Search the unrendered HTML for your price, product name or one-line definition. If it shows in a browser but not here, non-Google crawlers probably can't see it.
4. Check snippet controls. Look for noindex, nosnippet, max-snippet:0 and data-nosnippet. Templates often carry them by accident.
5. Ask the engines directly. Log out, turn memory off, and ask three buyer questions with search on and off. Note whether citations appear.

curl -s https://example.com/robots.txt
curl -s -o /dev/null -w "%{http_code}" -A "OAI-SearchBot" https://example.com/pricing
curl -s https://example.com/pricing | grep -ci "your price"
curl -s https://example.com/pricing | grep -ci "nosnippet"

Step 5 is the same idea as the search on vs off test.

Work through the stages in order. An earlier failure makes the later stages irrelevant.

• Robots.txt blocks a search bot. That engine's search can't show your page. OpenAI's bots page says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. Allowing GPTBot doesn't fix this, because each setting is independent.
• Robots.txt allows, but the request gets a 403. Your firewall or CDN is blocking the crawler. OpenAI's help centre says you need both: OAI-SearchBot allowed and OpenAI's published IP addresses let through.
• Facts missing from the raw HTML. Vercel found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. Server-render the facts that matter.
• Everything passes, but answers still skip you. Access is fine. The gap is in retrieval or citation, where engines often cite other sites. Being named on those sites is associated with getting into answers. See which sites AI cites most.
• An answer with no citations. It was probably answered from trained knowledge. Why ChatGPT answers sometimes have no sources explains this.

Be plain with yourself about what no simulator can see.

• The real queries are hidden. OpenAI's help centre says ChatGPT search typically rewrites your query into one or more targeted queries. Google's AI features may fan out across subtopics. A simulator tests the prompts you chose, not the queries the engine ran.
• Position inside the engine has no lever. C-SEO Bench (NeurIPS 2025) found where a document sits in the model's context mattered more than rewriting it. Nobody outside the engine can see or set that position.
• Content tweaks have weak evidence. A 2026 critical survey of 45 GEO studies found no technique with a stable, cross-platform causal effect on discoverability.
• Your test request isn't the crawler's. It comes from your IP, not the vendor's, so a firewall that trusts verified bots may treat the two differently.
• Answers vary. Whether an engine searches at all changes per query and per user setting.

Use a simulation to find blocks you can fix. Don't use it to predict a citation. More on this in can I simulate what AI search would say about my brand.

If you'd rather not run curl by hand, 99Visibility's free AI Search Simulator runs the same kind of model in your browser. Read its output the way this page reads any simulation. Treat access results as findings to fix and retrieval results as a model.

To see what engines actually say, the free checker on this site asks up to three buyer questions and scores the answers. It's a snapshot, not continuous tracking. The Claude search and citations guide covers why Claude's results can differ between runs.

Is a simulator the same as tracking AI visibility? No. A simulator models how an engine would handle a question. Tracking repeats real questions over time and records what engines actually answered. Neither sees the queries engines run internally, so both rely on prompts you chose. A simulator is most useful for finding access blocks. Tracking is more useful for seeing change over time.

Does allowing GPTBot help ChatGPT find my site? No. OpenAI documents GPTBot as a training crawler. OAI-SearchBot controls whether your site can appear in ChatGPT search answers. OpenAI says each setting is independent, so you can block GPTBot for training and still allow OAI-SearchBot for search. Check both separately in your robots.txt.

Why does Google behave differently from the other engines? Googlebot renders JavaScript, and Google's AI Overviews and AI Mode draw on the regular Google index. Google says a page must be indexed and eligible to show with a snippet to appear as a supporting link, with no extra technical requirements. There's no separate robots.txt token for these AI features.

Can a simulator tell me why a competitor gets cited instead of me? Only partly. It can show whether your pages pass the access stage. It can't show the engine's rewritten queries, its candidate list, or the order pages entered the model's context. Those decide much of which source gets cited, and none of it is observable from outside the engine.