AI quotes an old price of ours

What is actually happening

You changed your price. Your website shows the new number. Then a prospect pastes a screenshot of ChatGPT quoting the old one, with confidence, and sometimes with a citation to a page you forgot existed.

There is no single AI memory to correct. An answer about your price is assembled at the moment it is asked, from some mix of these:

• Live retrieval. The app fetches pages at question time and reads what is on them today.
• A stored index. The app searches an index built on an earlier crawl, so it reads your page as it looked on the crawl date.
• Model memory. The number was common enough in training text that the model recites it without fetching anything.
• A third party. A directory, marketplace, app store, affiliate roundup or comparison post carries the old figure, and that page is easier to fetch and parse than your own.

The fourth case is the one founders underestimate. If your own pricing page is behind a sign-up wall, loads its numbers with JavaScript, or says "contact us", then the cheapest readable price on the open web is whatever a blogger wrote in 2023. The engine is not being unfair. It is quoting the only parseable number it can see.

Who is doing the reading

Last verified: 17 October 2026

That is five apps and four vendor documentation pages you can open right now. Check your server logs and your robots.txt against that list before you blame the model.

One more primary source matters. Google documents a Product structured data format, which uses the schema.org Offer type with price, priceCurrency and priceValidUntil fields. A machine readable price with a validity date is the clearest signal you can give.

How to confirm it is happening to you

Screenshots from one prospect are an anecdote. Do this instead.

1. Write the question the way a buyer writes it, without your brand name. For example: "how much does project management software for a 10 person agency in India cost per month". Naming your business primes the engine to look you up, and the reading becomes worthless.
2. Write two or three brand named control questions such as "what does [brand] charge". These are diagnostic only. Keep them separate from the blind set.
3. Ask every question in all five apps on the same day: ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. They answer the same question differently on the same day because they read different sources.
4. Use a fresh chat for each question, logged out or in a clean profile, so memory and personalisation do not colour the answer.
5. Record four things per answer: the date, the app, the exact number quoted, and every citation link shown.
6. Repeat the set on at least three separate days. A single run tells you almost nothing. Report a range, not a percentage.
7. Hunt the source. Search the wrong figure as an exact quoted string in a normal search engine. Also search it with your brand name. The page that still carries it usually surfaces in the first few results.
8. Audit your own estate. Grep your site for the old number: pricing pages, landing pages, help centre articles, PDFs, sitemaps, old campaign URLs, and your structured data.

A worked example you can repeat

This is an illustration with invented numbers, not a customer.

A team raises its plan from INR 9,000 to INR 12,000 per month, excluding 18 percent GST, effective 1 September 2026. On 17 October 2026 they run eight blind questions in five apps, three days apart.

• Day one: two apps quote INR 9,000, two quote a vague "around INR 10,000", one declines to give a number.
• Exact string search for "INR 9,000" plus the brand returns three pages: an old /pricing-2025 URL that was never redirected, a partner directory listing, and a comparison blog post from March 2026.
• The team's current pricing page renders the number in JavaScript and has no Product structured data.

Fix list, in the order below: redirect /pricing-2025, add server rendered text plus an Offer with priceValidUntil, email the directory, and ask the blog for a dated correction. On the re-measure 30 days later, the two apps that fetch live pages quote INR 12,000 plus GST. The app that was reciting from memory still says INR 9,000. That split is normal and it tells you which lever you still have to pull.

What to change, in order

1. Your own pricing page, as visible text. Not an image, not a number injected after load. Include the currency, the billing period and the tax treatment in words, for example "per month, excluding 18 percent GST".
2. Add a date. Put "Prices effective DD Month YYYY" on the page, and keep a one line changelog of previous prices and their end dates. Dates let an engine prefer the newer fact.
3. Add structured data. Use Product markup with an Offer, and set priceValidUntil.
4. Kill the stale copies you own. 301 redirect old pricing URLs to the current one. Update or unpublish PDFs, decks hosted on your domain, and help centre articles. A 404 is better than a live old price.
5. Fix feeds and listings. Directories, marketplaces, app stores, review sites and affiliate pages. Each one needs its own update request. Keep a sheet of who you asked and when.
6. Check robots.txt. If you block the fetchers named in the table above, you have blocked the only route to your corrected number. Decide deliberately, not by default.
7. Get the correct number repeated off your own site. One corrected page of yours is weaker evidence than three independent pages agreeing. Ask for dated corrections, with the old figure struck and the new one stated.
8. Re-measure on a schedule with the same blind questions, so you can see movement rather than guess at it.

For the full audit sheet and test design, see #.

How long before it moves

Honestly: it depends on which of the four sources was feeding the answer, and we will not invent a number for you.

• Live retrieval answers change as soon as the fetched page changes, because nothing is stored.
• Index backed answers change after the next crawl of the corrected page. That is the vendor's schedule, not yours.
• Third party pages change when the third party edits them, which is a human timeline and often the slowest step.
• Model memory is the stubborn one. If the old price is baked into training text, it can persist across a measurement window and beyond. The practical counter is weight of fresh, dated, agreeing evidence.

Measure every 30 days and compare ranges, not single runs. If two readings have overlapping ranges, treat them as not having moved. That discipline stops you from celebrating noise.

Where AI Knows Us fits

AI Knows Us measures across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. Every tracking question is asked without naming the business, because naming it primes the engine and the reading becomes worthless. Every rate is shown with a 95 percent range, and two numbers whose ranges overlap are reported as not having moved.

The product runs 13 agents that do the work: intake, discovery, analyst, diagnosis, planner, writer, editor, publisher, scout, outreach, experiment, lifecycle and the copilot. So the fixing happens as well as the measuring: pages get written and published, and other sites get asked to name the business. AI Knows Us is made by Clyra Labs in India.

Paid plans are not open yet. The product takes no payment today and the free scan repeats every 30 days, which matches the re-measure cadence described above.

For how we compare the measurement tools in this space, see #.

Who should not choose AI Knows Us

• If you only need dashboards and alerts and have your own content team to do the fixing, tools such as Profound, Peec AI, Otterly.AI and Scrunch AI are worth evaluating on their own pages. We do not quote their prices here because prices change and we will not state a figure we have not checked on the day.
• If your problem is a retail or marketplace price feed, your first stop is the marketplace's own merchant support, not an AI visibility tool.
• If you need a signed contract, a procurement process and a named account manager today, note that paid plans are not open yet.
• If you want a single percentage score with no range attached, our reporting style will annoy you.

FAQs