The lead result
Publishing prices helps, but it doesn't guarantee an answer
Many business owners assume that if a price is on their website, an AI assistant can find it. We tested that assumption.
Before asking any AI engine anything, we sorted 1,123 US small businesses into three groups by where their prices could be found. For 647 of them, we verified prices on the business's own website. Then we asked Google's Gemini and Perplexity the same question about every business, once. The question asked what the business charges for a typical service in its category.
Gemini surfaced prices for 52.2% of the businesses with verified prices on their own site. Perplexity surfaced them for 48.4%. That is about twice the rate for businesses where we found no public prices. So publishing prices is associated with a much better chance of an AI answer, and the engines still failed to surface a price answer about half the time.
52.2% and 48.4%
How often Gemini and Perplexity surfaced prices that we had verified on the business's own website. For businesses with no prices we could find, the rates were 27.7% and 22.1%.
Pricing surfaced, by cohort and engine
Whiskers show 95% confidence intervals. An association, not a measured cause. Declined answers count as not surfaced.
| Where the business's prices are | Engine | Surfaced | 95% CI |
|---|---|---|---|
| On its own website (647 businesses) | Gemini | 52.2% | 48.4% to 56.1% |
| On its own website (647 businesses) | Perplexity | 48.4% | 44.5% to 52.2% |
| No public prices we could find (444 businesses) | Gemini | 27.7% | 23.7% to 32.0% |
| No public prices we could find (444 businesses) | Perplexity | 22.1% | 18.5% to 26.2% |
| Where the business's prices are | Businesses | Gemini surfaced prices | Perplexity surfaced prices |
|---|---|---|---|
| On its own website | 647 | 52.2%, 338 of 647 · 95% CI 48.4% to 56.1% | 48.4%, 313 of 647 · 95% CI 44.5% to 52.2% |
| Only on an outside platform | 32 | 12 of 32 | 13 of 32 |
| No public prices we could find | 444 | 27.7%, 123 of 444 · 95% CI 23.7% to 32.0% | 22.1%, 98 of 444 · 95% CI 18.5% to 26.2% |
CI is the 95% Wilson confidence interval. The outside-platform group has only 32 businesses, so we report it as counts, not rates. A declined answer counts as "not surfaced."
On Gemini, businesses with prices on their own site were surfaced 1.9 times as often as businesses with no prices we could find. On Perplexity the ratio was 2.2. This is an association, not a measured cause. We did not control for differences between the groups, such as the mix of categories and cities.
AI builds answers from sources you didn't publish
The engines didn't rely only on the business's own website. We counted every site each engine cited across all of its answers.
Perplexity's five most-cited sites were:
- yelp.com, 438 citations
- fresha.com, 214
- trustanalytica.org, 201
- tripadvisor.com, 165
- groupon.com, 163
Gemini drew on a different mix. Its most-cited sites were google.com with 151 citations, ubereats.com with 58, facebook.com with 36 and grubhub.com with 34. It also cited ordering and booking platforms that Perplexity's top sites didn't include: order.online and toasttab.com with 18 each, mindbodyonline.com with 17, square.site with 8 and zenoti.com with 5.
Most-cited sites, per engine
Perplexity: citations per site, Perplexity's own totals
Gemini: citations per site, Gemini's own totals
Each list is on its own axis, scaled to that engine's own totals. Bar lengths are not comparable across the two lists.
| Engine and site | Citations |
|---|---|
| Gemini: google.com | 151 |
| Gemini: ubereats.com | 58 |
| Gemini: facebook.com | 36 |
| Gemini: grubhub.com | 34 |
| Gemini: order.online | 18 |
| Gemini: toasttab.com | 18 |
| Gemini: mindbodyonline.com | 17 |
| Gemini: square.site | 8 |
| Gemini: zenoti.com | 5 |
| Perplexity: yelp.com | 438 |
| Perplexity: fresha.com | 214 |
| Perplexity: trustanalytica.org | 201 |
| Perplexity: tripadvisor.com | 165 |
| Perplexity: groupon.com | 163 |
None of Perplexity's 20 most-cited sites is on the list of 38 pricing sites we built from the businesses' own websites and the pages they link to. Fresha ranked second even though it never showed up as a pricing site when we crawled the businesses' own websites and the pages they link to.
The point isn't which engine preferred which site. An AI engine may build its answer about your business from information you didn't publish, don't control, and don't link to. Their citation behavior does not consistently privilege a business's own page over a third party's summary of it.
The 32 businesses whose prices live only on a booking or ordering platform show the same thing. For them, the platform page was a minority source, and the engines leaned on third-party summaries of the business. That group is small, so we treat it as an example of the mechanism, not a rate.
Three internets
What a person can see, what a conventional crawler can discover, and what an AI engine can retrieve and use are not always the same.
The 444 businesses in the "no public prices" group are a checked result. We crawled each business's website and the pages it links to, and human reviewers audited a random sample of 60 of these businesses. None of the 60 had pricing hidden on a platform we missed.
Even so, Gemini answered yes, with citations, for 123 of these businesses, or 27.7%. Perplexity did the same for 98, or 22.1%. That is roughly 1 in 5 to 1 in 4 businesses, depending on the engine.
Three internets (conceptual, not to scale)
Conceptual illustration, not to scale; circle sizes and overlaps are not measured.
| Engine | Surfaced pricing | Of businesses with no public pricing found |
|---|---|---|
| Gemini | 123 | 444 |
| Perplexity | 98 | 444 |
Two examples show how this happens:
- A restaurant's online menu ran on a Toast ordering page that our headless browser could never render. In an early test run, both engines said yes for it, each citing five sources.
- For a property management company, Perplexity cited ten listing and directory sites, including loopnet, zillow and apartments.com, and quoted $99, $75 and $149. Those prices exist on the web and are attached to the business, but on sites its own website never links to.
The 444 count is an accurate tally of businesses for which our crawl-and-audit process found no public pricing. It is not a count of businesses whose prices AI engines can't see. AI engines work over an information environment that doesn't map neatly onto a business's own site.
A yes is not the same as a correct price. Only 7 of these 444 businesses had a price our reviewers could quote, so we can't check most of the prices engines gave for this group.
AI visibility is not one stable answer
Both engines got the identical question about every business. They still differed in whether they answered, whether they found prices, which sources they cited, and sometimes which prices they gave.
Gemini declined to answer 286 of the 1,123 questions, or 25.5%. Perplexity declined 1. That gap drives much of the difference in the headline table. Counting only the questions Gemini answered, its rate rises to 64.9% for businesses with prices on their own site and 42.6% for businesses with no prices we could find. So the table does not show that one engine is better at finding prices. It shows two engines that behave differently under the same question. A declined answer still leaves the customer without a price.
Three businesses in the study had prices on several sites that didn't match each other. One was a Philadelphia med spa. Gemini said yes, citing the spa's website and its page on zenoti.com, and quoted ten prices from $200 to $800. On the same day, Perplexity said no while citing twelve sites, including the spa's own website, and quoted $200, $250 and $300. Three businesses is a case study, not a rate. When a business's prices disagree across sites, which price an AI answer carries depends on the engine, and the business can't predict it.
We checked prices only where our reviewers had quoted them. For the 77 on-site businesses with quoted prices, Gemini matched exactly 37 times, came within 20% 5 times and got it wrong 12 times. Perplexity's counts were 33, 8 and 16. The rest made no price claim. These counts cover a small subset, and the study gives no overall accuracy rate.
AI visibility did not behave like a fixed property of a website, or even of a business. It is an interaction between a business's public information environment and an engine's retrieval behavior. Either side can change, so the answer can change.
Why published prices don't always become answers
This section comes from an exploratory follow-up analysis of the same data. It was not part of the main study's checked figures, and its patterns need more rigorous retesting before anyone treats them as rules.
Most failures to surface an answer weren't discovery failures. Gemini said no for 183 businesses whose prices we had verified on their own site, and Perplexity for 334. In most of those answers the engine reached the business's website. Gemini cited the business's own site in 84% of its no answers and Perplexity in 91%. Only 21 of Gemini's no answers, 11%, and 18 of Perplexity's, 5%, looked like a true miss, with no price seen and the business's own site not cited. About half the no answers quoted prices anyway, 47% for Gemini and 51% for Perplexity.
The information didn't answer the question. In many cases the engine found prices that didn't match what was asked:
- teacher-training fees on a yoga studio's site, when the question was about a membership
- a $400-a-year dental plan, when the question was about the price of a single cleaning or crown
- per-unit asking rents at a property-management company, where the query wording itself proved ambiguous about what the company "charges"
- hair product prices, when the question was about salon services
- prices for a salon's location in another city
Often the engine read these correctly. A plan isn't a single-service price, and another city's prices don't answer a question about this one.
Presentation tracked with the result. Businesses whose sites had the fewest pricing signals in our crawl, 1 or 2, had prices surfaced 33% of the time on Gemini and 29% on Perplexity. Businesses with more than 10 signals reached 63% and 60%. Businesses with dollar amounts visible near the top of a subpage's text were surfaced 57% of the time on Gemini and 53% on Perplexity. Businesses with a pricing-style page address but no dollar amounts near the top of the text reached 44% and 42%. These are associations. Sites with many pricing signals probably also have dedicated price pages and clear service names, so this doesn't mean that repeating a price more often will get you found.
Presence is not the same as answerability. An answer passes through four steps:
- Retrieval. The engine finds the business's pages.
- Extraction. It reads the prices on them.
- Semantic matching. It decides whether those prices answer the question asked.
- Answer. It reports a price, or says it can't.
The answerability chain
Exploratory follow-up analysis
| Measure | Gemini | Perplexity |
|---|---|---|
| Own domain cited, share of no answers | 84% | 91% |
| True retrieval misses, share of no answers | 11% | 5% |
Answerability is the degree to which a business's public information can support a correct answer to a real customer question.
A business can pass the first two steps and still fail the last two. In this data, retrieval failures were comparatively rare. Many failures appear to occur after retrieval, at the semantic-matching step.
Why answerability matters more than a visibility score
Running more prompts across more engines and computing a score doesn't fix this. Take a hypothetical score like "63% AI visible." Its meaning depends on which engines were asked, how the prompts were worded, how refusals were counted, which retrieval path each engine took, and when the measurement ran. More measurements make that procedure more precise. They don't show that a stable underlying visibility state exists.
The answers are variable. What stays put is the public information they draw on: what exists about a business, where it lives, whether it is consistent, and whether it actually answers the questions customers ask. That is the more durable thing to measure.
Growfer audits that information environment. It looks for missing or conflicting information and shows a business where its representation to AI systems may break down.
What this means for your business
These are practical readings of associations, not proven causes. The study did not measure sales, bookings or customer behavior, and it doesn't show that any schema, markup or other technical change makes prices more visible to AI.
- Publish your prices on your own website. Businesses that did were surfaced about twice as often, though about half still weren't.
- Make each price answer the question a customer would ask. Say which service it covers, whether it is a single service, a plan or a membership, and which location it applies to.
- Keep your information the same everywhere it appears. When prices differ between your site, your booking page and review sites, which price an AI answer shows depends on the engine.
- Check what third-party sites say about you. In this study Perplexity cited Yelp more than any other site, and both engines drew on sources outside the business's own website.
- If your prices live only on a booking or ordering platform, don't assume that covers you. For those businesses, the platform page was a minority source in the engines' answers.
How we ran the study
Population. We started with 1,484 independent businesses with reachable websites, drawn from Foursquare Places listings with chains excluded. After qualification, 1,123 businesses made up the study. Businesses without a working website were excluded.
Cities. New York, Los Angeles, Chicago, Houston, Phoenix, Philadelphia, Miami, Atlanta, Denver and Seattle.
Categories. Ten business categories:
- hair salons and barbers
- med spas
- nail salons
- massage and day spas
- pet groomers
- auto detailing
- fitness, yoga and Pilates
- property management
- dentists
- restaurants
Sorting businesses by where their prices are. We crawled each business's website and the pages it links to, then sorted it into one of three groups: prices on its own website, 647 businesses; prices only on an outside platform, 32; or no public prices we could find, 444. Human reviewers audited 277 businesses. A re-check of 60 on-site businesses confirmed 59. A random audit of 60 no-prices businesses found hidden platform pricing at none of them, with a 95% confidence interval of 0% to 6.0%. We froze the three groups on September 22, 2026, before asking any AI engine about them. No engine answer was used to sort any business.
Engines. Google's Gemini, model gemini-2.5-flash grounded with Google Search, and Perplexity, model sonar with a medium search context size. We queried both through their APIs, not their consumer apps.
Dates. We asked every question on September 30, 2026. Gemini ran from 19:13 to 21:05 UTC and Perplexity from 21:06 to 22:05 UTC. The engines returned 2,646 answers with zero failures. The total cost of the engine questions was nominal, low enough that running a study like this is economical.
The question. Each engine got the same question about each business, once, with no follow-ups. The service named in the question came from a fixed list, one service per category.
What does [business name] ([city]) charge for [service]? Answer only from [website] or pages that clearly host its pricing; cite the URLs you rely on.
The engine was asked to end with a verdict of yes, no or refused. A refused answer counts as a verdict and was never re-asked.
Price accuracy. Where our reviewers had quoted a business's prices, we compared the engine's prices to them. A price within 20% is labeled close and is not counted as a match.
Discovery questions. We also asked each engine two general questions per city and category, 200 per engine, such as which local businesses offer a service and what they charge. Businesses from the study came up with pricing context 92 times in Gemini's answers and 43 times in Perplexity's. That asks whether an engine picks a business out of every business it could recommend, which is a different question from whether it can find a known business's prices. Low counts there are not evidence of poor visibility, so we don't report them as visibility.
Not measured. ChatGPT and Claude were not measured. The study was designed for four engines, and every result on this page covers only the two we measured.
Limitations
- Two engines on one day. The results are a snapshot from September 30, 2026, not a model of how the engines behave over time.
- ChatGPT and Claude were not measured, so these results may not hold for them.
- We used the engines' APIs. Answers in their consumer apps may differ, and we didn't test those.
- The 1.9 and 2.2 ratios are unadjusted associations. The groups differ in their mix of categories and cities, and we didn't control for that.
- Cities and categories aren't evenly represented, and some city and category combinations have few businesses. The analysis pools them.
- The outside-platform group has only 32 businesses, so we report it as counts. It also undercounts platform pricing in general, because a headless browser can't capture some ordering pages.
- Gemini declined 25.5% of questions. Counting those as "not surfaced" lowers its rates, and the answered-only figures above show by how much.
- We could check price accuracy only where reviewers had quoted prices. That was 77 of 647 on-site businesses, all 32 platform businesses and 7 of 444 no-prices businesses. There is no overall accuracy rate.
- Each business got one question about one service chosen for its category. Some businesses don't sell that service in that form, and a few had a category or city label that didn't match their website.
- The follow-up analysis is exploratory. It uses simple text checks, and its pricing-density pattern can't separate how much pricing a site shows from how well that pricing matches the question.
- We did not measure revenue, conversions, bookings or customer behavior.
Sources
- Growfer research, "The Split-Pricing Study: Final Report," revision 6, October 5, 2026. The full study report, with complete method, derived tables and limitations.
- Growfer research, "On-site pricing not surfaced as an answer," companion analysis, revision 2, October 7, 2026. The exploratory follow-up analysis.
- Foursquare Places API. The source of the business listings used to build the study population.
- Google Gemini API, gemini-2.5-flash grounded with Google Search, and the Perplexity API, sonar. The two engines measured, queried September 30, 2026.
