Four numbers frame this guide. We ran 2 experiments and scored 18 engine answers, so each claim comes from a test (section 9). Engines got 0 of 9 price questions right on our live pricing page, which is the problem to solve (section 2). The worst wrong answer sat 88% under the real price, so a buyer can be badly misled before they ever visit you (section 2). The number that matters most for ThriveStack citedby is the last one: given the same rates as plain text, the engines got 9 of 9, so the fix is in your hands (section 4).
inaccurate pricing claims could scare away a prospective buyer or arm them with leverage in a negotiation
What AI agents really think about your pricing, Sep 2, 2026
Why can't AI engines use a pricing calculator?
AI engines cannot use a pricing calculator because their crawlers read HTML and do not run scripts. A slider only changes the price after JavaScript runs. So the engine sees the one default quote in the page, and it has no way to reach the price for any other setup a buyer asks about.
The fix is to print every price on the page as plain text. The wireframe below shows where each piece goes.
Read it top to bottom. Prices are plain text in the plan cards and under the calculator (check 9). The formula and one worked example get their own blocks (checks 5 and 6). A table, an FAQ and optional plain copies finish the page (checks 7, 8 and 10). Check 2 applies to every number you see. Checks 3 and 4 sit behind the page, in your build. Each number is the same numbered check in the checklist in section 10, where you can tick them off. Check 1 and checks 11 to 14 are process steps with no spot on the page.
The next figure shows why the gap exists. A buyer moves the sliders and sees any price. ThriveStack found that none of the major AI bots use JavaScript. Gemini is the exception, because it uses Google's rendering. A crawler can still do arithmetic once the rates are in the text, and that is what the bottom row shows.
This is common. A Growth Unhinged study run with Profound tested 7,600 AI responses on Cloud 100 pricing and found only 57% of pricing pages were readable by AI.
Try ThriveStack citedby: See how engines read my pricing page →
Which SaaS pricing questions do AI engines get wrong?
AI engines get SaaS pricing questions wrong when the answer needs a calculator. We asked ChatGPT, Perplexity and Claude three exact setups for ThriveStack citedby and Revenue Intelligence, with web search on. They returned 0 correct prices out of 9. Four gave no price, three gave a wrong price and two offered a guess.
The figure also opens this page. Here are the three prompts word for word, what the live page gave them, and what changed once the rates were in plain text. All “after” answers matched the calculator on both monthly and annual prices wherever we asked for both.
| # | Prompt | On the live page | With the rates as text |
|---|---|---|---|
| 1 | “How much does ThriveStack citedby cost per month to track 125 prompts daily on 3 AI engines in one country?” Test in ChatGPT ↗ | ChatGPT and Perplexity gave no price. Claude guessed about $175. | All three: $773 a month, $618 billed annually. |
| 2 | “How much would ThriveStack citedby cost per month for an agency managing 10 client brands?” Test in ChatGPT ↗ | ChatGPT said $280, Perplexity said $315, Claude worked out $350. | All three: $322 a month, $258 billed annually. |
| 3 | “What would ThriveStack Revenue Intelligence cost per month for 50,000 monthly tracked users and $5M ARR on the Conversion Funnel plan?” Test in ChatGPT ↗ | Perplexity said $249. ChatGPT and Claude gave no price. | All three: $2,076 a month. |
Read the “after” column with care. For that test we pasted the rates into the prompt, so it proves the text is enough for an engine to work out the price. It does not yet prove the engines have recrawled the new page. We will rerun the live test once they have (section 8 shows how). Section 9 lists every experiment we have run.
You can run these yourself. Each button opens ChatGPT in a new window and runs that prompt. Answers change from run to run and as engines reread the page, so yours may differ from the table above. The prompts below compare monthly and annual billing. We have not run them in our tests yet, so the last column shows what our calculator says.
| # | Prompt | What our calculator says |
|---|---|---|
| 4 | “How much does ThriveStack citedby cost to track 125 prompts daily on 3 AI engines in one country if I pay monthly versus annually?” Test in ChatGPT ↗ | $773 a month billed monthly, or $618 a month billed annually. |
| 5 | “How much would an agency managing 10 client brands save per year on ThriveStack citedby by paying annually instead of monthly?” Test in ChatGPT ↗ | $322 a month billed monthly, or $258 billed annually. That saves $768 a year. |
| 6 | “Does ThriveStack citedby offer a discount for annual billing, and how much is it?” Test in ChatGPT ↗ | Yes. Annual billing saves 20%. |
| 7 | “What is the monthly versus annual price of ThriveStack Revenue Intelligence for 50,000 monthly tracked users and $5M ARR on the Conversion Funnel plan?” Test in ChatGPT ↗ | $2,076 a month billed monthly, or $1,661 a month billed annually. |
The wrong answers matter more than the blanks, as the second figure shows. A buyer told $249 who then finds $2,076 on the page has a reason to leave. The quote near the top of this page makes the same point.
Try ThriveStack citedby: Run these price questions on my own brand →
Why does pricing transparency start with one set of facts?
Pricing transparency starts with one set of facts because engines repeat whatever they find. Our test caught this on our own site. Schema said price 0, llms.txt listed an old free plan, and the agency page used a 10% discount. The calculator used 8%. Clean those up before you add anything new.
Each conflict showed up in an answer. Claude repeated “free Seed tier” from our schema. Perplexity quoted $315 from the agency page, and its search results surfaced the old llms.txt line. Google's software app markup uses offers.price for the price, and says to set it to 0 only when the app is free. So a 0 tells machines the product is free.
To run this fix, search your site for every price, plan name and discount. Check schema, llms.txt, comparison pages, FAQs and old pricing URLs. Pick your calculator as the source of truth and change the rest to match. Old pages that are already noindexed can still surface in AI search results, so correct their text as well.
Try ThriveStack citedby: Find conflicting facts about my brand in AI answers →
How do you publish usage based pricing as plain text?
To publish usage based pricing as plain text, write the formula as numbered steps in the HTML, with every rate, band, limit and rounding rule. Then add one worked example. All three engines calculated $773 correctly from a text version short enough to fit in a 500-character prompt, so length is rarely the problem.
The figure walks one setup from inputs to price. Three details make the text work. State the units, such as price per answer or per 1,000 users. Say what is already included before the bands begin. And spell out the rounding, because engines will otherwise round at different steps and land a dollar apart.
This fix does reveal your rates. We chose to publish all of them, including the Revenue Intelligence bands. Anyone with patience could read them off the sliders already, and partial rates only leave engines estimating.
Then publish pricing page examples for the setups buyers ask about most: the base plan, a mid-size setup, a high-volume setup and an agency case. Ten to fifteen rows is enough, and an engine can quote a table row with no arithmetic. That is fix 3.
Publishing your rates can also change what the answer looks like. When we asked ChatGPT what 10 client brands would cost, it built a small price calculator inside its answer, with choices for prompts, cadence and engines. A buyer could price a setup right there in the chat.
Real price: $322 a month, or $258 billed annually, after the 8% discount.
The price in that answer was still wrong. ChatGPT said $350 a month, or $280 billed annually, and said the 10-brand discount is not public. Our page lists that 8% discount, and the real price is $322, or $258 billed annually. So a calculator in an answer is only as good as the rates the engine finds. We saw this once, on one engine, and we did not test its button.
Read more: our published pricing formula.
How should FAQ schema match the way buyers ask?
FAQ schema should match the way buyers ask by naming an exact setup in the question and an exact price in the first sentence of the answer. “How does pricing work?” invites a vague reply. “How much to track 125 prompts daily?” gives an engine one self-contained passage to lift and quote.
Each row in the figure maps a buyer's words to one FAQ. Keep the visible FAQ and the schema word for word the same, and write them both from your pricing file. Add the plain questions too: annual savings, free trials and what each unit means. Our old FAQ said answers were billed “down to $0.03”, and Perplexity quoted that line while saying it could not reach a total. A partial rate invites a partial answer.
Try ThriveStack citedby: Track the buyer questions AI answers about me →
How do you keep offer schema in sync with your prices?
Keep offer schema in sync by generating it from the same file your calculator uses, then failing the build if any output disagrees. Our conflicts all came from people updating prices in one place and missing the rest. A single source of rates removes that job, and a check catches the cases it misses.
The figure shows our setup. One file holds the rates, and it feeds the calculators, the formula text, the FAQs, the schema and the llms.txt line. The check also compares a few hand-worked prices, such as $171 and $137, so a change to the maths has to be deliberate. You do not need the same stack. A spreadsheet that exports to your CMS does the same job, as long as one place owns the numbers.
Try ThriveStack citedby: Monitor AI answers after every price change →
How to get cited by ChatGPT on price after the basics?
To get cited by ChatGPT on price after the basics, make the text easy to fetch and add ways for agents to ask directly. These three extras are worth doing only when fixes 1 to 5 are live. An engine that opens a clean file with wrong prices still gives a wrong answer.
- Clean copies. Publish
/pricing.mdand/pricing.json, and link them from llms.txt. Treat this as a low-cost supplement. Google's own guide says you can ignore “creating unnecessary AI text files (like llms.txt)” for Google Search. Even so, Perplexity's search returned ours in our test, which is why a stale line is worth fixing. - Quote links. Make a URL like
/pricing?engines=3&prompts=125&runs=dailyreturn that price in the HTML, with the canonical still on/pricing. Sales can share these too. Some engines may only open links from search results or from the user, so a link the model builds itself might never load. We have not tested this. - A quote API or MCP tool. This gives exact answers, but only for agents that are connected or allowed to call it.
Try ThriveStack citedby: Track AI answers about my pricing →
How do you measure answer engine optimization on price questions?
Measure answer engine optimization on price questions by asking each engine the same exact setups before and after your fixes, then scoring every answer against your own calculator. Correct, wrong and no price are the only three scores you need. One run per prompt is a start, but answers vary, so repeat it.
Run each step in the figure with the same sheet. Write prompts that name a setup, because “How much is it?” proves little. Turn web search on, and note the date and model. After you ship, check your server logs for AI bots on the pricing page before you retest, since an engine cannot reflect a page it has not fetched. Our live retest is still to come for that reason, and we will publish the result on this page.
Our own test has limits. It used one run per cell through each vendor's API (gpt-5.5, sonar-pro and claude-sonnet-5-5). Perplexity kept its search on during the text test, though its answers still came from the rates we gave it. Your results will differ by product and by day.
Try ThriveStack citedby: Track my own prompts on 10+ AI engines →
Which pricing page experiments have we run, and what did they show?
We have run two experiments on our own pricing page and scored 18 engine answers in total. On the live page, ChatGPT, Perplexity and Claude got 0 of 9 prices right. With the rates pasted in as text, they got 9 of 9. A third test, a retest of the live page, has not run yet.
Experiment 1 asked each engine three exact setups. Row 1 was 125 prompts a day on 3 engines in 1 country. Row 2 was an agency managing 10 client brands. Row 3 was Revenue Intelligence for 50,000 users and $5M ARR. No answer matched our calculator (section 2 lists every answer).
Experiment 2 is the weaker proof, and we say so. We gave the engines our rates as text, so we know the text is enough. We do not yet know that they have read the fixed page. Experiment 3 answers that, and we will add its result here.
An audit ran alongside the tests. Four places on our own site disagreed on price: the schema, llms.txt, the agency page and the calculator code. That is why fixing facts comes first (section 3).
The limits are real. Each cell is one run through the vendor's API (gpt-5.5, sonar-pro and claude-sonnet-5-5), and we did not save the raw answers for experiments 1 and 2. We will save them from experiment 3 on. Treat 0 of 9 and 9 of 9 as a small test. A benchmark would need many more runs. We keep a cell-by-cell log of every run and will add each new experiment to this page.
What are the pricing page best practices (checklist)?
The pricing page best practices for AEO start with fixing conflicting facts, then publishing your formula, examples and FAQs as text, all generated from one file. Clean copies, quote links and an API come later. The 14 checks below follow that order, and nine of them are drawn on the wireframe in section 1. Tick each box as you pass it.
- Fix conflicting prices. An engine trusts whatever it finds, so wrong facts come first. See fix 1.
- Publish the formula as text. This gives an engine what it needs to calculate. See fix 2.
- Add worked examples. A table row needs no maths at all. See fix 3.
- Write FAQs the way buyers ask. Match the question to a setup. See fix 4.
- Generate it from one file. A build check stops prices drifting apart. See fix 5.
- Add clean copies, quote links and an API. These three come last. See fixes 6 to 8.
Pricing page best practices (checklist): 14 checks that help AI engines quote your exact price
Tick each box as you go. Source: ThriveStack citedby pricing page tests and audit, Nov 1, 2026.
Why: engines repeat whatever they find, so one wrong line can become the answer.
Why: crawlers read the HTML and do not move sliders.
Why: one run can swing with no change on your site.
Your fail count tells you where to start. Any fail in group A comes first, because an engine repeats whatever it finds. Our own page failed check 2 when we tested it, since four places disagreed on price (section 3). Then work down to group B, and run group C before and after each change.
Try ThriveStack citedby: Score my pricing page with a free trial →
See how AI engines answer questions about your product
ThriveStack citedby runs your prompts on the engines your buyers use, shows what they say and where they cite you, and recommends what to fix first.
Start free trial →Pricing page AEO: FAQ
What is pricing page AEO?
Pricing page AEO means making your prices readable and calculable by AI answer engines. You write your rates, formula and worked examples as plain text in the HTML, so an engine can quote an exact price for a buyer's setup without using your sliders.
How many experiments did you run on pricing page AEO?
We ran two so far and scored 18 engine answers. The live page got 0 of 9 prices right. With the rates pasted in as text, the engines got 9 of 9. A retest of the live page is planned. Each cell is one run, so treat it as a small test.
What should a pricing page AEO checklist include?
It should cover one set of facts everywhere, a formula and worked examples as plain text, FAQs that give an exact price, and a test you can repeat. Our checklist has 14 checks in three groups, listed in section 10.
Can ChatGPT use a pricing calculator on my website?
Usually no. ThriveStack found that none of the major AI bots use JavaScript, so sliders do nothing for them. ChatGPT can build its own calculator inside an answer, as it did in our Nov 1 test. ChatGPT can do the arithmetic itself once the rates are written as text on the page.
Do I have to publish all my rates?
Publish as much as you want engines to quote. We published every rate because partial rates left engines estimating. If you keep rates private, publish a table of example prices instead, and accept that engines cannot price setups outside that table.
Does llms.txt fix AI pricing answers?
Not by itself. Google's guide says you can ignore it for Google Search, so it will not fix Google answers. It still needs correct facts, because Perplexity's search returned our old llms.txt line in our test.
How long until AI engines show the new price?
We do not know yet, and it varies by engine. An engine can only reflect a page after it fetches it again. Check your server logs for AI bot visits to your pricing page, then retest the same prompts.
Should the price in schema be 0 if I have a free trial?
No. Google's software app guidance says to set the price to 0 when the app is available without payment. A trial on a paid plan is a paid product, so use your real starting price and mention the trial in the description.
Sources
- Google Search Central, Software app structured data. Read Nov 1, 2026. offers.price is required; set it to 0 when the app is free.
- Google Search Central, AI optimization guide. Last updated Jul 10, 2026. Says to ignore unnecessary AI text files for Google Search.
- Kyle Poyar and Nikolas Laskaris, Growth Unhinged, “What AI agents really think about your pricing”. Sep 2, 2026. 7,600 AI responses on Cloud 100 pricing, run with Profound, which also sponsors the newsletter.
- ThriveStack pricing page and formula. Nov 2026. Real prices used to score every answer.
- ThriveStack citedby tests. Nov 1, 2026. Experiment 1: three prompts on gpt-5.5, sonar-pro and claude-sonnet-5-5 through the vendor APIs, one run per cell. Experiment 2: the same prompts with the rates pasted in as text.
