What Is Prompt Tracking for AI Search?
Prompt tracking is running the same fixed list of buyer questions through AI assistants on a regular schedule and recording which brands are named or cited. It exists because AI answers vary: in a Yext study of local "top 10 near" prompts, four AI engines named the same top local business in only 4% of cases 3. One run proves little; a repeated set shows a trend.
- Prompt tracking = a fixed list of prompts, the same assistants, run on a schedule, with every answer logged.
- Answers differ by engine. Yext found all four engines named the same top local business in only 4% of 491,972 cases 3.
- Citation sources can swing within weeks. ChatGPT's Reddit citations fell from close to 60% to about 10% of tracked responses in six weeks 4.
- Freeze the list for a month at least 1. Compare weeks, not single runs.
What is prompt tracking?
Prompt tracking is a measurement routine. You write down questions your buyers would ask an AI assistant. You run them in the same assistants on a set day. You record which brands appear and which pages are cited.
A prompt here just means the question or request typed into the assistant. The word "tracking" means you repeat the exercise often enough to see change.
It is the AI-search equivalent of rank tracking. Classic rank tracking follows a page's position for a keyword. Prompt tracking follows whether you appear in an answer, because there is no fixed position to follow.
How does prompt tracking work?
A tracker holds a fixed prompt list and checks it repeatedly. Semrush's Prompt Tracking checks a fixed set of prompts daily against one chosen platform, records whether your brand appears, and requires the prompt set to stay constant for at least a month or trend lines break 1.
It starts the day you set it up. Semrush says its tracking has no backfill, and that a true 28-day comparison takes about eight weeks 1. The same logic applies to a spreadsheet: you cannot recover last month's answers afterwards.
Paid tools automate the running and counting. A spreadsheet does the same work by hand, more slowly and with fewer prompts.
Why does a single run mislead?
AI answers are not stable lists. They vary between engines, between days and between runs of the same prompt. Ahrefs puts it plainly: AI answers are inherently volatile, so the way to measure visibility is to track a sample of prompts 5.
Engines disagree. Yext asked Gemini, OpenAI, Perplexity and Anthropic the same local question for 491,972 shared cases between February and August 2026. All four named the same top business in 4% of cases, and four different ones in 48% 3. Yext sells local listings software, and the prompts were one fixed format.
Time matters too. Semrush tracked more than 230,000 prompts across three AI platforms for 13 weeks in 2025. ChatGPT's citations of Reddit fell from close to 60% of responses in early August to about 10% by mid-September. The drop was isolated to ChatGPT in that study, while Perplexity and Google AI Mode moved much less 4. Semrush could not say why. If you had run your prompts once in August, you would have drawn the wrong conclusion about September.
How often should you run each prompt?
Weekly is enough for a small team doing it by hand. For software, Profound found one run a day gave nearly the same visibility as ten, and says portfolio construction matters more than extra runs for citation share.
Profound compared one run a day with ten a day over 14 days in June 2026, with 753 prompts on seven platforms. Visibility was 78.7% against 80.4%, a gap of about 2 points 2. This is a vendor study of a single window, so treat it as a hint. The takeaway is that more prompts help more than more repeats of one prompt.
For a manual log, the practical rule is simple. Run every prompt in every chosen assistant once a week, on the same day, and keep at least four weeks of history before drawing a conclusion.
What does a good tracking row look like?
Each row should hold enough to re-score the answer later. Store the full answer text, not only a yes or no, because your counting rules will change.
The example below is a template with made-up data. Copy the columns, not the values.
- Week | Prompt | Assistant | You named | You cited | Others named | Notes
- W1 | best payroll software for 10 staff | ChatGPT | N | N | Competitor A, Competitor B | answer saved in /logs/w1-chatgpt-01.txt
- W1 | best payroll software for 10 staff | Perplexity | Y | Y | Competitor A | cited /pricing
- W2 | best payroll software for 10 staff | ChatGPT | N | N | Competitor A, Competitor C | new competitor
How do you choose which prompts to track?
Choose prompts a buyer would type, not keywords a marketer would target. Spread them across the buying journey: learning (what is), comparing (best, versus) and deciding (price, reviews, near me).
Start with about 30. Semrush suggests that number as a starting point 1. Keep branded prompts, which include your name, in a separate list, because they flatter you. The query fan-out explorer can show how an assistant splits a question, which gives ideas for related prompts, and how to build a prompt list for tracking walks through it.
- Write each prompt once, in final wording, and do not edit it.
- Include the same competitors in every comparison prompt.
- Add a column for the city or country if answers are local.
- Only add new prompts as a new list, so old trends stay valid.
How do you tell a real change from noise?
Use a stricter test than your instinct. Our rule of thumb, not a statistical standard: count a change as real only if it holds in three weekly checks in a row and shows up in at least two assistants.
Citations also churn on their own. Profound followed 883,000 pages for a year and found the median page lost half its citation share 11 days after its peak 6. A drop after a good week is therefore normal, and not proof that something broke.
Write down every change you make to the site, with its date. Without that note you cannot connect a movement to an action.
When is prompt tracking not worth the effort?
Skip it while the basics are missing. If AI crawlers cannot reach your site, or you have no pages that answer buyer questions, tracking will report absence every week and teach you little. Check crawling with the AI crawler checker first.
Also skip paid tools if you only need a baseline. A spreadsheet and an hour a week is enough to learn whether you are visible at all. Pay for software when the manual routine takes too long, or when you need many prompts, many assistants or stakeholder reports.
Frequently asked questions
What is the difference between prompt tracking and rank tracking?
Rank tracking follows a page's position for a keyword on a results page. Prompt tracking follows whether your brand is named or cited in an AI answer to a fixed question. AI answers have no stable position, so presence over repeated runs replaces rank.
How many prompts should I track?
Begin with about 30 real buyer questions. Semrush suggests that figure and notes its plans track 25 to 200 prompts 1. A small, well-chosen list that you keep fixed is more useful than a long list you keep changing.
Should I track ChatGPT, Gemini or Perplexity?
Track the assistants your buyers use. Engines often disagree: in one study four of them named the same top local business in only 4% of cases 3. Start with two or three, and add more only if you can keep the routine going.
Can I do prompt tracking for free?
Yes. Keep a spreadsheet with the prompt, assistant, date, whether you were named or cited, and who else appeared. Run it weekly. It is slower and covers fewer prompts than software, but it works. See the tracking spreadsheet guide.
Sources
- How to build an AI visibility dashboard in Semrush · SemrushOpened October 20, 2026. Supports: Prompt Tracking checks a fixed set of prompts daily on one platform; keep prompts constant for at least a month; no backfill; about eight weeks for a true 28-day comparison; start around 30 prompts; plans track 25 to 200. Published 12 October 2026.
- Is once a day enough? · ProfoundOpened October 20, 2026. Supports: 753 prompts, 7 platforms, 1 to 14 June 2026: visibility 78.7% at one run a day vs 80.4% at ten; citation share more sensitive to prompt mix than run count; vendor study of one window.
- Four AIs, one question, four different answers · Yext ResearchOpened October 20, 2026. Supports: 491,972 shared local cases, February to August 2026: all four engines named the same top business in 4% of cases and four different ones in 48%; unbranded 'top 10 near' prompts; vendor of local listings software.
- The Most-Cited Domains in AI: A 3-Month Study · SemrushOpened October 20, 2026. Supports: 230,000+ prompts, 13 weeks from 14 July 2025: ChatGPT's Reddit citations fell from close to 60% of responses to around 10% by mid-September; isolated to ChatGPT; cause not established.
- How to measure AI search visibility · AhrefsOpened October 20, 2026. Supports: AI answers are inherently volatile; the way to measure visibility is to track a sample of prompts.
- The half-life of an AI citation is 11 days · ProfoundOpened October 20, 2026. Supports: 883,000 pages across seven answer engines over one year: median page lost half its citation share 11 days after peak; vendor study.