ChatGPT Visibility Tracker: What to Measure and What to Pay

A ChatGPT visibility tracker is software that asks ChatGPT a fixed set of buyer prompts on a schedule and records, for each answer, whether your company is mentioned, recommended, compared or cited. It measures what an answer says, not a search position, so the useful output is a rate across repeated runs.

ChatGPT can search the web when a question benefits from current information, and answers that use search can include inline citations or a Sources panel. That is why the response and its cited sources should be saved together (OpenAI, ChatGPT search).

We use "first-appearance order" deliberately. A generated answer is not a fixed results page, so calling the second company listed "position two" creates false precision.

One answer is a single observation from a system that generates text probabilistically, so it cannot separate a lasting pattern from ordinary variation. In the SparkToro and Gumshoe.ai study, 600 volunteers ran 12 prompts 2,961 times across ChatGPT, Claude and Google's AI in November and December 2025. Two responses almost never matched.

The same study shows what does hold steady: how often a brand appears. Across 994 responses to a headphones prompt, Bose, Sony, Sennheiser and Apple each appeared in 55% to 77% of them, while a prompt about brand design agencies, a crowded field, gave the top names visibility in the 30s and 40s. One cancer center appeared in 69 of 71 ChatGPT responses (97%) yet came first in only 25 of those (SparkToro and Gumshoe.ai study).

That is why we report a mention rate across many runs and treat any single "rank in ChatGPT" number as false precision. Run from a documented session with memory turned off and record the market, so personal settings do not become a hidden variable.

Spotting a 10-point move in a 40% mention rate takes about 185 runs per period, a 15-point move about 82, and a 20-point move about 47. Judging one rate is cheaper: about 30 runs of a prompt puts a single mention rate within roughly 17 points either way, and about 93 runs gets it within 10.

These figures are our model, not a vendor's. For a single rate we used the Wilson score interval, a standard way to put a margin of error around a share measured from a small number of runs, at 95% confidence (z = 1.96). It assumes the runs are independent. Each row uses the observed rate shown (the 1 of 3 row is an illustrative count); the 40% assumption applies only to the run-count formulas below the table.

For one rate at plus or minus 10 points and a 40% rate, the runs needed are 1.96² × 0.4 × 0.6 ÷ 0.10² = 92.2, rounded up to 93. For plus or minus 5 points in the worst case (a 50% rate), the figure is 385.

Comparing two periods is harder, because both rates carry noise. The margin on the difference is 1.96 × √(2 × 0.4 × 0.6 ÷ n) for n runs in each period. At 30 runs per period that is 0.248, about 25 points. At 60 runs per period it is about 18 points. To detect a difference d, the runs needed per period are 1.96² × 2 × 0.4 × 0.6 ÷ d². That gives 184.4 (rounded up to 185) for 10 points, 82 for 15 points and 47 for 20 points.

Our rule: report a change only when it exceeds the margin on a difference, which is about 1.4 times (the square root of 2) a single rate's margin. Then add a confidence rule. Investigate a movement that appears once, watch one that appears inconsistently, and report a directional change only when the same fixed prompt group moves across two reporting cycles and the move exceeds that margin.

Pooling helps. Twenty prompts run three times each gives 60 runs, and a pooled rate of 24 mentions in 60 carries the 29% to 53% range above. Treat that pooled rate as a portfolio average, because prompts differ in difficulty and no single prompt has earned that precision.

Always report the denominator. "Mentioned in 8 of 20 prompts across 60 runs" is informative. "Visibility improved" is not, unless the portfolio, run count, and test conditions are clear. Cadence sets your sample: a weekly schedule gives about four runs per prompt per engine each month (52 weeks ÷ 12), and a daily schedule gives about 30.

Run buyer questions in six types (brand, category, problem, comparison, alternative and purchase intent) on every engine your buyers use, not only ChatGPT. Start with 20 high-value prompts. Tag each one by prompt type, buyer role, use case, and priority so the final report can distinguish a category gap from a single weak comparison.

The examples assume a SaaS (subscription software) company.

Brand prompts put your name in the question, so they test accuracy, not discovery. Discovery is judged on the category, problem, comparison and alternative prompts where your name is absent. Take the wording from sales calls, support tickets and win-loss notes (sales notes on why deals were won or lost), and see Prompt Research for how to choose which prompts earn a place.

Use one canonical wording for each buyer job. Add two phrasing variations only when the intent remains the same and wording could reasonably alter the answer. Lock that list for a reporting period, then log additions, removals, and wording changes before the next cycle. This discipline matters because a changing portfolio can manufacture a gain or loss. If you replace difficult prompts with easier ones, the mention rate may rise even though your visibility has not changed.

Engines differ in what they cite. In our Citation Research Study 007 (fieldwork 5 October 2026), we asked 150 decision questions from India of ChatGPT with search on, Perplexity, Gemini and Google AI Mode, one run per question and engine. Google AI Mode cited a community platform (Reddit, YouTube or Quora) in 117 of 150 answers. ChatGPT did in 3 of 150. That is one run per question from one market, so read it as a direction, not a US benchmark. Our inference is that one blended score across engines hides this kind of gap, so each engine should be reported separately.

Every answer should be saved as a row with the exact prompt, run date and time, market, model, whether web search was on, the full answer text, and any cited URLs. A rate without those conditions cannot be repeated, and a screenshot is not a record.

Mark the first occurrence of your company, then label the description as correct, incomplete, or incorrect. Note whether the answer recommends the company, merely lists it, or frames it as unsuitable for the buyer's stated need.

A spreadsheet is a legitimate way to start. Twenty prompts run three times on one engine is 60 answers a month. The same 20 prompts daily on four engines is 20 × 4 × 30 = 2,400 answers a month, which is where hand-running breaks down and runs get skipped. Our free AI Visibility Audit gives a one-off check: paste a domain and in about 30 seconds, with no signup or credit card, it checks what AI engines say about your brand.

As of October 2026, monthly-billed plans on the providers' own price pages run from $29 to $1,499 a month, or about $1.22 to $7.50 per tracked prompt. Cost per prompt is a rough guide only, because engines, cadence and bundled tools differ. PageLens.ai publishes this page and appears in the table.

Per prompt is the monthly-billed price divided by tracked prompts, our arithmetic. Semrush plans also include keyword and site tools, so that column overstates its tracking cost. Otterly's four engines apply to all its plans, and extra engines cost more as add-ons. Annual billing lowers the monthly rate, by about 8% on our plans (PageLens.ai pricing) and 15% on Otterly's.

Cost per answer is a fairer test of cadence. Our Launch plan's 300 answers a week is 100 prompts on three engines, about 1,300 answers a month for $299, or $0.23 each. Growth gives 500 a day, about 15,000 over 30 days, roughly $0.05 each. Enterprise gives 1,400 a day, about 42,000 over 30 days, roughly $0.04 each. Our plans also include prompt research and 25, 50 or 100 content pieces a month on Launch, Growth and Enterprise.

Seven checks decide whether a tracker's numbers deserve trust. Each one has a pass and a fail you can verify in a demo or a trial, including ours.

Run the hand check on a trial: ask 20 non-branded prompts three times each, then compare your rates with the vendor's. The AI Visibility Tool Trial Checklist turns this into a 14-day plan.

Route each finding to an owner with a due date: a missing mention goes to a content owner, an incorrect description to the page that should be authoritative, and a weak recommendation to clearer fit and limits. The goal is not to force a mention but to make public information clearer, more accurate and more useful.

• Missing Mention: Create or improve a page that directly answers the buyer job.
• Incorrect Description: Update the most authoritative owned page with specific, current language.
• Weak Recommendation: Clarify fit, limitations, and use cases rather than adding vague claims.
• Unhelpful Citation Mix: Review the cited topics and build genuinely useful supporting content.

OpenAI says any public website can appear in ChatGPT Search, recommends allowing its search crawler for summaries and snippets, and does not guarantee top placement (OpenAI publishers and developers FAQ). Crawl access and content quality are prerequisites, not promises.

We also separate answer visibility from referral traffic. A click does not prove that a company was recommended, and a recommendation can influence a buyer without generating a click. Track AI Traffic in GA4 covers the visit side in Google Analytics 4.

We run a fixed set of buyer prompts on a schedule across the engines on your plan and keep the answers, mentions, sentiment and sources, so you can read what each engine said, not only a score. Launch tracks ChatGPT, Google AI Mode and Perplexity weekly. Growth adds Gemini and Grok and runs daily. Enterprise adds Claude and Copilot, for seven engines. All plans include unlimited users.

Our prompt research reads your site and category and surfaces the comparison, alternative, best-of and fit questions buyers ask. It groups them by funnel stage and scores each by how winnable it is. We recommend a daily plan when you need to confirm a change within a month, because daily runs give about 30 samples per prompt where weekly gives about four.

Prices for each plan are on our pricing page. To check your own prompt count, engines and run cadence against the seven checks above, Book a demo.