AI Visibility Tracking: How It Works

Every AI visibility tracker does roughly the same four things.

1. Build a prompt set. You pick the buyer questions that matter, such as "best payroll software for a 20-person company". This is prompt tracking in its plain form.
2. Ask each engine. The tracker sends each prompt to ChatGPT, Perplexity, Gemini, Claude, Google AI Overviews or whatever engines it covers. Some use vendor APIs. Others capture what a search results page shows.
3. Log the answer. It records whether your brand is named, which URLs are cited, which competitors appear, and sometimes whether the facts stated about you are right.
4. Repeat on a schedule. The same prompts run again next week or next month. The differences become the trend line.

Some APIs expose more than citations. OpenAI's web search tool returns every URL the model consulted separately from the URLs it cited, and its docs say the consulted list is often longer. A tracker that logs both sees which of your pages were retrieved and then dropped, not just which made it into the answer.

What comes out is a table of answers per prompt, per engine, per date. Everything else, including any score, is built on that table.

A tracker measures answers to prompts it chose. That is useful, but it is not the same as measuring what real users see.

The real queries are hidden

OpenAI's help centre says ChatGPT search typically rewrites your query into one or more targeted queries before sending them to search providers. Google says AI Mode may issue multiple related searches across subtopics. So the prompt a tracker types is not the query that retrieves pages. As of October 2026, Bing Webmaster Tools is the only surface that shows the grounding queries an AI actually ran, and only in its interface.

Search is optional per answer

Retrieval fires at the engine's discretion. ChatGPT decides per query whether to search. Claude.ai search is a user toggle. Published estimates of how often ChatGPT searches range from 18% (Profound) to 80.5% (cloro), because they measure different things on different populations. An answer without citations may come from the model's frozen training data, which nothing you publish this quarter can reach.

Samples undercount

In Otterly's 2026 experiment, its prompt-based tracker logged 3,934 citations for one test site, while Microsoft's own AI Performance report credited that site with 42,817. The time windows may not match, so it is not an exact ratio. It does show that any prompt sample sees a slice.

No rank exists

Google's Search Console report for AI features shows impressions only: no clicks, no click-through rate, no position. No ground truth exists for an "AI rank", so a tracker can report presence, absence and citation counts, not a position.

What this means: treat tracked numbers as a consistent proxy. They are good for spotting change and comparing engines. They are not a census of your AI exposure.

A trend line assumes the thing being measured stayed the same. In AI search it often doesn't.

• "ChatGPT" is not one model. OpenAI documents a February 16, 2026 knowledge cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path behind chat-latest. If routing or a plan default shifts, what ChatGPT knows about you can shift by months with no change on your side.
• Apps change defaults. As of October 2026, Google's release notes show Gemini 3.6 Flash selectable in the Gemini app since July 21, 2026, while Google AI Mode has defaulted to Gemini 3.5 Flash since May 19, 2026. Google also says AI Overviews and AI Mode "may use different models and techniques", so one Google line hides two products.
• APIs retire. Perplexity retired Sonar Chat Completions on September 27, 2026. On Perplexity's Agent API, web search is opt-in, so a tracker that migrated without enabling it gets ungrounded answers. The chart would show citations collapsing for a reason entirely inside the tool.
• Your CDN can change for you. From September 15, 2026, Cloudflare blocks Training and Agent crawlers by default on ad-displaying pages for new domains, new sites on existing accounts and existing free-tier customers. A site can lose some AI access without anyone editing anything.
• Your own index status can change. In Otterly's test, Google deindexed both AI-generated sites with no manual action, and Google's AI surfaces stopped citing them. Bing kept them, and Copilot kept citing them.

What this means: a step change in a trend line is a question, not a verdict. Line it up against known platform dates before you read it as a content effect. For the cross-model view, see LLM visibility tracking across models.

Start with the shape of the change, then look for the cause that fits.

Three reading rules help:

• Read each engine on its own line. In Otterly's experiment, Microsoft Copilot produced 2,718 of one site's 3,934 tracked citations (69%), while ChatGPT, Gemini, AI Overviews and AI Mode almost never cited it. A blended score would have hidden that.
• Annotate the chart. Mark every known model launch, default change, API migration and site deploy. Unexplained steps are the ones worth investigating.
• Compare like with like. Brand size dominates citation patterns. Benchmark against brands of similar size, or you are measuring size, not visibility.

Tracking tells you what changed. Only some of the causes are yours to fix.

In your control

• Rendering. Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. Content that appears only after JavaScript runs may be invisible to them. Server-render the facts that matter.
• Per-bot robots.txt. OpenAI says each crawler setting is independent. OAI-SearchBot controls ChatGPT search. GPTBot governs training only. Blocking the wrong one removes you from search answers.
• Network reachability. OpenAI says eligibility needs both robots.txt permission and a host or CDN that lets its search crawler through. A firewall can block what robots.txt allows.
• Index and snippet controls. noindex, nosnippet, max-snippet:0 and data-nosnippet keep content out of AI answers. Templates sometimes apply them by accident.
• Stated facts. A specific price, number or one-line definition on a fetchable page gives an engine a sentence it can quote.

Partly in your control

• Third-party mentions. When engines cite other sites instead of yours, being named on those sites is associated with appearing in the answer. No effect size is promised.

Not in your control

• Which model an app routes to, whether it searches for a given answer, how it rewrites the query, and where your page lands in its context. A critical survey of 45 GEO studies (arXiv, July 2026) found no technique with a stable, cross-platform causal effect on discoverability.

Set tracking up so the numbers can survive scrutiny.

1. Fix the prompt set. Choose a small set of buyer questions and freeze it. Changing prompts mid-series breaks the comparison as surely as a model swap.
2. Control the session. ChatGPT may use saved memories when rewriting a search query. Run logged out or with memory off, and record it. The reproducible ChatGPT prompt tracking method covers this in detail.
3. Record the conditions. For each run, log the engine, the model or plan where known, whether search was on, and whether citations appeared.
4. Check access before the first run. Confirm your robots.txt lets OAI-SearchBot, PerplexityBot and Claude-SearchBot in, and that your firewall does too. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
5. Take a baseline. The free checker on this site gives a snapshot for up to three buyer questions. It is a starting point, not continuous tracking.
6. Pick a cadence and keep it. Re-run on a fixed schedule so changes line up with dates. See how often to re-run AI visibility checks.
7. Annotate and investigate. When a line moves, use the table above. Fix access and factual errors first, because those are the causes you can verify and remove.

• Trusting one blended score. Engines use different models, cite in different units and fire search at different rates. Any single score averages answers with and without live retrieval.
• Reporting a rank. No engine exposes position data for AI answers. Report presence and citations.
• Treating a citation as influence. Citation counts are a noisy proxy. Retrieved pages outnumber cited ones, and research on how models cite shows citation output can change without the sources changing.
• Reading every drop as a content problem. Check platform changes and access first. Rewriting pages for AI has no support: C-SEO Bench found most such rewrites ineffective or harmful.
• Assuming robots.txt is the whole gate. It is necessary, not sufficient. Test the firewall too.
• Allowing GPTBot to get into ChatGPT search. GPTBot governs training. Search uses OAI-SearchBot.
• Taking a vendor's prompt sample as the full picture. It is a slice of an input nobody outside the engine can see.