Why AI Quotes an Old Price: A Dated Price-Audit Test Across Your Website, Feeds, and AI Answers

An assistant quotes an old price of yours for one of three reasons, and they need different fixes: the current price is not in the bytes your server returns, so it was taken from somewhere else; two public sources of yours disagree, so the assistant picked one; or the answer was built from memory with no live search at all. The number that tells you how bad your own situation is, is the count of tested AI price answers that match your current dated source price, divided by all price answers tested, with the assistants named and the test date printed. We have not run that match test, on our own domain or any client's, so this page publishes no match rate. As of 29 September 2026 the figure does not exist in our records and we will not estimate it. What we hold instead are four dated price failures measured on clawlaw.in in legal technology in India, including the one that explains most of them: on 6 August 2026 the pricing page rendered its prices only after scripts ran, so what a crawler received contained no prices at all, and the prices the assistants quoted had come from an app store listing instead of the company's own site.

That is the whole mechanism in one sentence. The assistant was not being careless. It answered a price question from the only price it could read. Everything below is built so you can find out which of the three reasons applies to you.

Define the current price and the terms attached to it

Before you can score any answer as right or wrong you need one row that is the truth, and most businesses cannot produce it on request. Write it down with the following nine fields, for every product or plan you sell.

• Plan name, exactly as a buyer would see it, with no internal codenames.
• Amount and currency, written in full.
• Billing period, monthly or annual, and whether the figure shown is the monthly equivalent of an annual commitment.
• Tax treatment. Whether the figure includes GST. This single field causes more scoring disputes than any other, because an assistant quoting your price without tax and a page showing it with tax look like a mismatch and are not one.
• Unit. Per user, per seat, per matter, per company, per search. A correct amount attached to the wrong unit is a wrong price.
• Minimum commitment, including any minimum seat count.
• What is excluded, named individually: onboarding, support tiers, data migration, overage.
• Effective from date, and the date you last reviewed the row.
• The canonical URL where that price is published.

Put the effective date and the review date on the public page too, not only in your sheet. There is a measured reason. On 17 September 2026 Claude examined two clawlaw.in comparison pages that stated competitors' prices with no link, no date and no source. The figures turned out to be correct when the assistant checked them independently, and it still treated the pages as advocacy and used official sources instead. A correct price with no date beside it does not earn trust. That is also why this page does not restate a single competitor's price. If you want a competitor's number, open that company's own pricing page and record the amount, the date you read it and the URL together, in one row.

Enumerate every public price source

Name them. Do not count them. The following eleven are the places we would check for a business selling software or a service in India, and the list is the point, because the source you forget is the one the assistant reads.

• Your pricing page, as served, not as rendered in your browser.
• Any other page on your site that mentions money, including the home page, a plans comparison, a blog post from two years ago, and an FAQ.
• Your structured data, if your pages carry product or offer markup with a price in it.
• Your app store listings, on each store separately, including in app purchase tiers.
• Your marketplace listings, for example a cloud marketplace or a software directory that carries plan prices.
• Directories and review sites that publish a price field, each named individually in your sheet.
• Your own PDF material that is publicly downloadable: a rate card, a brochure, a proposal template that got indexed.
• Press coverage and interviews where a founder said a number out loud.
• Partner and reseller pages that publish their own version of your pricing.
• Any feed you supply, product feed, affiliate feed, or an API that returns prices.
• Archived copies of your old pricing page, which are public and which an assistant can reach.

For each one record the URL, the price it currently shows, the plan name it uses, and the date you checked. You now have a contradiction map. On 6 August 2026 that exercise on clawlaw.in found the website and the app store listing carrying different plan names and different prices for the same product, and ChatGPT noticed the contradiction and said so in its answer. An assistant that finds your sources disagreeing will often report the disagreement, which is worse for a buyer than either price alone.

Run the price consistency test

This is the protocol. It has two halves: what your own sources say, and what the assistants say. Run the first before the second, or you will not be able to score anything.

One. Fetch your priced pages the way a crawler does. Request each URL without running scripts and read what comes back. If the number is absent from that response, no amount of publishing will help, because the price is not in the document. This is the first test to run and the one most likely to explain everything.

Two. Freeze a prompt set of price questions. Between twelve and thirty prompts, saved to a file with a version number and a date. Write them in the words a buyer uses: how much does this cost, what are the plans, is there a free tier, what does the cheapest plan include, how does its price compare with the alternatives.

Three. Split the set into named and blind. Prompts that name your product measure whether your price is stated correctly. Prompts that do not name it measure whether your price is stated at all when somebody is choosing. Keep them in separate files, because they answer different questions and mixing them corrupts both rates. The reason is measured: on 27 July 2026 two runs of the same 78 questions happened on one day for clawlaw.in, and the run whose wrapper named the brand came back ranking it first on almost every question, while the blind run put the company second by breadth and absent altogether from the questions it most wanted to win. The first run was discarded.

Four. Fix and record the conditions. Assistant and mode, session state, location setting, language, device, timestamp. Price answers vary by country more than most answers do, because currency and tax handling change.

Five. Set the repeat count and print the arithmetic. Three fresh sessions per prompt per assistant. With twenty prompts, three assistants and three repeats the denominator is twenty times three times three, which is 180 recorded answers per round. Publish that multiplication next to the rate, every time, so a reader can see how the denominator was built rather than take it on trust.

Six. Score each answer against your canonical row with one of five verdicts. Exact match. Correct amount with wrong unit or period. Correct amount with wrong tax treatment. Outdated amount, meaning it matches a price you used to charge. No price given. Those five are not interchangeable and collapsing them into right and wrong hides the fix.

Seven. Record the source of the price in the answer, in two columns. Claimed sources, being whatever the assistant tells you. Verified sources, being the URLs you opened yourself. Treat the first as a claim. In the aiknowsus.com audit of September 2026 Perplexity withdrew its own earlier statement about having searched for each question, saying it could not honestly substantiate it, and in another batch said its claim to have searched all five was not adequately supported. That happened in three separate batches. An assistant's account of where it got your price is evidence of nothing until you open the link.

Eight. Record whether a live search happened at all, as a separate field, because it changes what the answer means. On 27 July 2026 ChatGPT confirmed it had run no live web search for any of 78 questions, which makes every price in those answers a memory of a price rather than a reading of one.

Results: mismatches by source and product

We publish no price match rate on this page, because we have not run the test. The cell is empty and stays empty until the protocol above has been completed and the captures published, at which point it will carry a numerator, a denominator, the assistants named and the date range, in one sentence.

What we do hold are four dated price findings from the clawlaw.in programme in legal technology in India. They are the nearest real evidence we have, they are not a match rate, and they are not about your sector.

• Prices absent from what a crawler received. 6 August 2026, Claude. The pricing page rendered prices only after scripts ran, so the crawler's copy contained no prices at all, and the prices the assistants quoted had come from an app store listing instead.
• Two public price lists in disagreement. 6 August 2026, ChatGPT. The website and the app store listing carried different plan names and different prices for the same product, and the assistant reported the contradiction in its answer.
• A correct but undated competitor price discounted. 17 September 2026, Claude. Two comparison pages stated competitors' prices with no link, no date and no source. The figures were correct when the assistant checked them independently, and it treated the pages as advocacy and used official sources instead.
• Price transparency driving the ordering. 17 September 2026, Claude. On a question about finding every case against a company, two enterprise vendors were ranked above clawlaw.in specifically because they publish explicit coverage lists, and the assistant said its ordering reflected price transparency and source authority rather than product quality.

Read the last one carefully if you are tempted to hide your pricing. The ordering in that answer was not about which product was better. It was about which vendor could be described in specific terms.

Correct the source, then notify or recrawl where supported

Work in this order, because the later steps do nothing if the first is skipped.

One. Put the price in the served HTML. Plain text in the document that comes back from a plain request. If your pricing is generated by a script, add a server rendered version of the same numbers.

Two. Make one page canonical and point everything at it. One URL that carries the full price table, the effective date, the tax treatment and the exclusions. Every other mention of money on your site links to it.

Three. Fix every contradicting source in your enumeration, one row at a time, and note the date you corrected each. App stores and marketplaces have their own review queues and will not update the day you ask.

Four. Remove or clearly date old prices rather than deleting the pages silently. A post from two years ago that says a plan costs a certain amount should carry a visible note saying it was accurate on that date and linking to the canonical page.

Five. Ask for a recrawl where the platform offers one. Search consoles, app stores and some directories provide a way to request a refresh of a page or a listing. Open the current documentation for each platform on the day you do this and follow it. We are not restating those instructions here with a date attached, because they are edited and a stale paraphrase of a submission process becomes a false instruction.

Six. Check crawler access while you are there. A price that is published and blocked is not published. Fetch each priced URL as each named crawler and record the status code and whether the price appears in the body.

Retest and track correction time

Rerun the frozen prompt set, unchanged, on a schedule, and log every answer with the same fields. Report the match rate per round with the arithmetic beside it, and report a second number that is more useful than the rate: the days between the date you corrected a source and the first round in which an answer quoted the new price on that assistant.

Do not expect a stable interval and do not let anybody sell you one. No assistant publishes a refresh schedule for a page it has already read. The nearest timing we hold comes from publication rather than correction: on 6 August 2026 ChatGPT cited a specific clawlaw.in article fourteen days after that page was published, in a zone where the same question set had named the company nowhere at baseline. That is one page, one assistant, one date, and it is not a refresh time.

Limits: no guaranteed refresh time across assistants

Six limits. Each one is a conclusion somebody has drawn wrongly from a test like this.

• No refresh time is published by any assistant, so any specific promise about how fast a price correction propagates is invented.
• We hold no price match rate, for our own domain or anybody else's, as of 29 September 2026.
• The four findings above are legal technology in India, on clawlaw.in, on ChatGPT and Claude. They are not a benchmark for retail, for manufacturing, or for any other sector, and moving them onto another sector's page would be the same error this page warns about.
• A match rate is a fact about your prompt set and your dates, not a property of the assistant.
• Tax and currency handling will produce false mismatches unless your five verdicts separate them, which is why the scoring has five options and not two.
• A price appearing correctly after a fix is not proof the fix caused it. Without a set of prompts you did not act on, you cannot separate your change from the assistant's own change that month.

Sources and change log

The findings above come from the clawlaw.in programme of July to September 2026, recorded in GEO_BASELINE_RESULTS_2026-07-27.md and the assistant audit files of 17 September 2026, and from the aiknowsus.com audit of September 2026 across 24 batches and 72 conversations. Of the four price findings, the crawler reading, the two price lists and the undated competitor price discounting can be checked from outside today. The counts taken over our capture folders cannot, until we publish the captures.

Version: 29 September 2026, first publication. The page will be updated when we complete a price match test and can print a numerator and a denominator, when any figure is corrected, and when the captures are published.

Common questions

Why would an assistant use an app store price instead of our own pricing page?

Because the app store page had prices in it and the website's response did not. That is exactly what happened on clawlaw.in on 6 August 2026, with Claude and ChatGPT on the same date. Fetch your pricing page without running scripts and read the response. If the numbers are not there, this is your answer and nothing else needs investigating first.

We do not publish prices at all. Is that safer?

It is not neutral. On 17 September 2026 Claude ranked two enterprise vendors above clawlaw.in on a coverage question and said its ordering reflected price transparency and source authority rather than product quality. A business that cannot be described in money terms is harder to recommend specifically. If you cannot publish a number, publish a structure: what the pricing depends on, a range with the conditions named, and a date.

Should we put our competitors' prices on our comparison pages?

Only with a link, a date and the source named for each figure, and even then expect them to be checked. On 17 September 2026 two clawlaw.in comparison pages carried correct competitor prices with none of those three things, and Claude treated the pages as advocacy and went to official sources instead. The safer instruction, and the one this page follows, is to name the pricing page the reader should open and tell them to record the amount, the date and the URL together.

How many price prompts are enough?

Twelve is enough to find out whether your price is readable and whether your sources contradict each other. It is not enough to publish a rate and defend it in a meeting. Whatever number you choose, print the multiplication that produced your denominator beside the rate, every single time.

An assistant told us where it got our price. Can we act on that?

Open the URL first. Record what the assistant said in a claimed sources column and what you verified in a separate column. In the aiknowsus.com audit of September 2026, Perplexity withdrew its own claim to have run a live search for each question across three separate batches, saying it could not honestly substantiate it. Self reported sourcing is a lead, not evidence.

What to do first

Fetch your pricing page today without running scripts and check whether your prices are in the response. Then list every other public place your prices appear, with the amount and the date you read it, and find the contradictions. Those two steps cost an afternoon and they resolve most old price answers before you have run a single prompt.