Can I get historical AI visibility data, or does tracking only start at signup? What we tell clients
What we tell clients about historical AI visibility data: tracking starts at signup, but analytics, Search Console and logs add context. Here is how we build baselines.
New clients ask this in the first call: can we see what ChatGPT said about us last year? We understand why. A before-and-after story needs a "before". The honest answer is that for the thing most people mean, the answers to your prompts, tracking starts when you start. Here is what we explain, and what we do about it.
Why there is no backfill for AI answers
AI answers are generated on request. Unless somebody asked your exact prompt on a given day and saved the response, nothing exists to retrieve. A platform can run your prompts now and store every response from then on. That is how Promptwatch works: it runs your prompt panels daily per model, country and language, starting when the prompts are created. We have found nothing published suggesting any platform can reconstruct what a model would have said months ago, and we would be skeptical of anyone who claims it.
So we say plainly to clients: your history starts the day we set up tracking, and every week of delay is a week you will not get back.
What we use to fill the gap
We do not leave the client with nothing. We build context from places that do keep history.
Web analytics. We look at AI referral traffic for as far back as retention allows, filtering for sources like chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com. This shows visits that clicked through. It does not show mentions without a click, and it is imperfect, since referrers can be missing, but it gives a sense of direction.
Search Console and Bing Webmaster Tools. Search Console keeps about 16 months. We export it early. It is traditional search data, but it shows whether organic demand and rankings moved around the time the client suspects a change. Promptwatch's Search Insights overlays Search Console and Bing data on AI crawls and citations, so once we are tracking, we see them together.
Server and CDN logs. If the client keeps raw logs, we search for historical hits from GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and the rest. Retention is often short. We have not seen documentation that Promptwatch imports historical logs, so we treat its crawler data as starting at connection. If your CDN keeps logs, export what you can before connecting.
Change history. The CMS revision log, release notes and, where useful, the Wayback Machine. They let us explain a visibility shift after the fact.
Pre-collected datasets are different
Some products sell a pre-collected index. Ahrefs Brand Radar's AI Visibility Index is one, covering more than 496 million prompts modeled from keyword data, retested monthly on a 90-day window. That is useful for market-level research and for seeing how a category looks, but it is not your own prompt history, and it is modeled from keywords, not from your buyers' actual conversational questions. We treat it as directional.
How we build a baseline fast
- Day one. Write 30 to 50 prompts across topics, prompt types and personas, including competitor comparisons. Set the Brand Book so the tool counts mentions correctly.
- Day one. Connect crawler logs and install the visitor script, so attribution starts too.
- Weeks one and two. Change nothing on the site. Let the tool collect.
- Week three. Review the baseline: visibility score, share of voice, position in answer, sentiment, cited pages.
- Week four. Start changes, one at a time, with dates logged.
Promptwatch's A/B period comparison and prompt trends are built for this approach. For clients who need to start with no budget, the free Explore tier gives 10 prompts on ChatGPT only, and the no-login AI Visibility Report gives a snapshot.
Keep the prompt set stable
The most common way teams ruin their own history is by rewriting prompts every month. If you edit prompts, the trend line stops being comparable. Add new prompts, tag them with a date, and leave the core set alone.
What to tell your boss
Say this: we cannot recover the past, but we can measure from today, and we will pair it with analytics and search history to show direction. Then show a four-week baseline and the first change.
Promptwatch is where we start that clock for clients. The sooner it starts, the sooner a trend line exists.