All answers were collected on 24 September 2026 and coded against vendors’ public pricing evidence as it stood on that date. Pricing pages and marketplace listings change, so the figures describe that snapshot, not the vendors’ current prices.
Every figure on this page is a CoreAEX study finding from one dated sample of 120 answers. Findings are described with their counts, denominators and scope, and are not tagged. Recommendation marks the practical steps that follow from those findings. The study did not test whether any of those steps changes what an answer engine says.
The short answer
CoreAEX asked four AI search surfaces the same pricing question about 30 B2B SaaS products and coded all 120 answers against each vendor’s public pricing evidence on 24 September 2026.
Among the 80 answers about products with fully or partly sales-gated offers, 62 (77.5%) preserved the fact that the relevant offer had no public price. They identified it as Contact Sales or quote-based and did not attach an unsupported number to it.
That did not make the whole answer accurate. In the fully undisclosed group, only 14 of 40 answers were accurate overall. The problem was often not an invented official price. It was the structure around the price: third-party estimates, wrong or retired tiers, billing mistakes, minimums and allowances, and separate products merged into one.
Contact Sales pricing does not automatically make an AI answer wrong. It creates an evidence gap. Most responses in this study preserved the absence of an official price, but many filled the remaining space with third-party estimates, incorrect plan structures or unsupported commercial details.
What we tested
We selected 30 B2B SaaS products on purpose, 10 in each of three pricing groups (“strata”):
- UNDISCLOSED: At sampling, no current numeric price for the core product appeared in the frozen official public pricing evidence. The official route required contacting sales or requesting a quote.
- PARTIAL: At least one current official numeric price was public, while at least one named plan, tier or material product configuration required contacting sales.
- PUBLIC_COMPARATOR: The standard named offers relevant to the prompt had current official numeric prices. A separate bespoke enterprise route could exist, but it did not define the product’s ordinary pricing state.
The equal groups were a design choice. They let us compare pricing situations side by side. They do not reflect how common each situation is among SaaS vendors, and none of the percentages below is a market estimate.
| UNDISCLOSED | PARTIAL | PUBLIC_COMPARATOR |
|---|---|---|
| Gong Revenue AI | Slack | Basecamp |
| ServiceNow ITSM | Asana | Pipedrive CRM |
| Vanta Trust Management Platform | Airtable | Freshdesk |
| Drata GRC Platform | Miro | Help Scout |
| Ironclad Contract Lifecycle Management | Linear | Buffer |
| WalkMe Digital Adoption Platform | Webflow | Tally |
| Highspot AI Sales Enablement Platform | Zapier | Fathom Analytics |
| Outreach Agentic Revenue Workflow Platform | Contentful Platform | Ghost(Pro) |
| Workday Adaptive Planning | Semrush Local Toolkit | Transistor podcast hosting |
| Greenhouse Recruiting | Sentry | Simple Analytics |
Each product was put to four surfaces: ChatGPT Search, Google AI Mode, Microsoft Copilot and Perplexity. That gave 120 runs, one per product on each surface. All 120 produced a valid response; none was void and none returned no answer.
Every run used the same fixed prompt, with only the product and vendor names changed. It asked for current plans and prices, the billing basis, important usage or seat limits, and supporting sources. Two independent coding passes then classified each answer, and every disagreement was settled against the captured answer and the evidence. The full prompt and conditions are in the method notes.
Most answers preserved Contact Sales, but that was not enough
The main measure was whether an answer preserved the absence of a public price. An answer passed when it correctly identified the relevant offer as Contact Sales or quote-based and did not assign an unsupported numeric price to that offer.
An answer could still pass if it included a third-party estimate, as long as it clearly labelled the estimate and kept it separate from the vendor’s official offer. Whether that estimate was accurate, attributed or supported was coded separately. This measure applies only to the 80 UNDISCLOSED and PARTIAL answers, because the comparator products had public standard pricing.
| Absence preserved | Answers | Share of 80 |
|---|---|---|
| Yes | 62 | 77.5% |
| Partly | 4 | 5.0% |
| No | 14 | 17.5% |
The 77.5% counts only full passes. The other 18 answers, 22.5%, did not preserve the absence fully. That figure combines four partial passes and 14 failures, so it is not a count of invented prices. A partial pass meant the line between an outside estimate and the vendor’s official offer was blurred, or other unsupported pricing language weakened it.
The marketplace check
Vanta, Drata and Highspot were placed in UNDISCLOSED because their own pricing pages routed buyers to sales. During coding, some answers surfaced AWS Marketplace listings, sold by those vendors, with numeric prices for limited configurations. Those listings are not complete direct-sale price cards, and they are not ordinary self-serve pricing. But they do mean that “no public price” depends on which route and configuration a statement is about.
We kept the original groups so the design did not change after collection. We then reran the main measure without the 12 answers about those three products:
| Absence preserved | Answers | Share of 68 |
|---|---|---|
| Yes | 53 | 77.9% |
| Partly | 4 | 5.9% |
| No | 11 | 16.2% |
The strict rate moved from 77.5% to 77.9%. These three mixed cases do not drive the main result.
Preserving the boundary is not the same as an accurate answer
The UNDISCLOSED group shows the gap most clearly. Twenty-seven of those 40 answers preserved the unavailable price. Only 14 were accurate overall. The two figures measure different things. The first asks whether the answer kept the sales-gated offer free of an unsupported number. The second asks whether every material pricing claim in the answer held up: plan names, billing basis, minimums, allowances and any figures quoted.
An answer could say “Contact Sales” correctly and still get the plan structure wrong. That contrast is the core of this study.
Across all 120 answers, the overall results were:
| Result | Answers | Share of 120 |
|---|---|---|
| Accurate | 59 | 49.2% |
| Partly accurate | 43 | 35.8% |
| Inaccurate | 18 | 15.0% |
| Unverifiable | 0 | 0.0% |
Fewer than half were accurate on every material pricing claim. That does not mean the other 61 were wrong. Most of them, 43, were partly accurate: they often got the main pricing situation right and missed or misstated a narrower detail. Eighteen were inaccurate overall.
| Pricing stratum | Accurate | Partly accurate | Inaccurate | Total |
|---|---|---|---|---|
| UNDISCLOSED | 14 | 16 | 10 | 40 |
| PARTIAL | 21 | 15 | 4 | 40 |
| PUBLIC_COMPARATOR | 24 | 12 | 4 | 40 |
In this sample, full accuracy was less common in the UNDISCLOSED stratum than in the other two. That is an observed association. The study was not designed to show that withholding a price caused the difference.
What filled the pricing gap
Where no official number was available, many answers included someone else’s.
Third-party numeric figures appeared in 47 of 120 answers. They appeared most often in the UNDISCLOSED stratum: 28 of 40 UNDISCLOSED answers, compared with 11 of 40 PARTIAL and 8 of 40 PUBLIC_COMPARATOR answers.
Of those 47 answers:
- 23 attributed the figure to a third party;
- 11 mixed official and third-party figures; and
- 13 gave a third-party figure without attributing it.
Attribution helps a reader see where a number came from. It does not make the number accurate. A clearly labelled estimate can still be old, based on a different company size or attached to the wrong plan.
Two further coded measures:
- Unsupported numeric prices. Twenty-four of 120 answers contained a price the public evidence did not support. The measure applied to 109 answers, so the rate among those was 24 of 109, or 22.0%. Eight more were coded partly for unsupported-price issues. “Unsupported” means the public evidence did not support the number. The study did not establish how each number came to be in the answer.
- Missing qualifications. Sixteen of 120 answers left out an important condition, such as an annual-billing requirement, a minimum commitment, a boundary between product families or the fact that a number was only a benchmark. These omissions were coded separately from false statements.
Patterns we saw
The coded cases showed several recurring patterns. They describe this sample; they are not a complete list of how pricing answers go wrong.
- Third-party ranges beside a quote-only offer. Sometimes clearly labelled, sometimes placed so they looked like the vendor’s own tier price.
- Buyer segments turned into plans. Company-size bands or external benchmarks became named “current tiers” the vendor did not offer.
- Retired plans presented as current.
- Billing basis changed or dropped. Annual-equivalent prices shown as ordinary monthly prices, or per-user and flat-rate structures mixed together.
- Separate products merged into one price ladder.
- Correct headline, wrong detail. The answer said “Contact Sales” correctly, then added unsupported minimums, implementation fees, allowances or contract terms.
An example
Slack, ChatGPT Search. The answer attached a $45 figure to Enterprise+, although Slack’s official route for that plan was Contact Sales. The rest of the Slack pricing in the answer was largely correct. Because that one number was placed directly on the sales-gated offer, the answer was coded inaccurate overall. One material number can change the result.
If you are trying to work out why a specific price is missing or stale in answers about your own product, our diagnostic guide to missing and stale SaaS prices covers the checks in order.
Public pricing did not eliminate errors
Publishing prices did not guarantee accurate answers. Twenty-four of the 40 PUBLIC_COMPARATOR answers were accurate overall. The rest were partly accurate or inaccurate.
Eight of 120 answers presented an outdated price when checked against the dated official evidence, and seven of those eight were in the PUBLIC_COMPARATOR stratum. A public price can still be misreported when an answer draws on a retired plan structure or an older comparison page.
Basecamp, Google AI Mode and Microsoft Copilot. Both answers presented Basecamp’s retired per-user pricing as current.
Simple Analytics, Perplexity and Google AI Mode. Both answers made two separate errors. They presented the retired Starter and Business tiers as current, and they attached a number to the bespoke Enterprise route. Both were coded inaccurate.
The Simple Analytics case also shows that a “public” product can still have a custom route. PUBLIC_COMPARATOR means standard plans were publicly priced, not that every configuration was.
A citation is not the same as support
Most answers showed a source. Fewer showed a source that fully backed the price claim beside it.
| Measure | Answers | Denominator | Rate |
|---|---|---|---|
| Visible citation supplied | 91 | 120 | 75.8% |
| Official source visibly cited | 68 | 120 | 56.7% |
| Citation fully supported the price claim | 49 | 87 applicable | 56.3% |
| Citation partly supported the price claim | 35 | 87 applicable | 40.2% |
| Citation did not support the price claim | 3 | 87 applicable | 3.4% |
Citation support was coded for 87 applicable answers. In 49 of them, the cited page fully supported the claim. In 35, it supported part of it. A cited page could be the right vendor’s pricing page and still not support the specific plan, amount, billing period or unit the answer gave.
One interface limitation affects the first row. Twenty-eight ChatGPT Search answers kept only internal citation markers when copied, so they were coded as “no visible citation supplied.” Two ChatGPT Search answers kept visible official links. This records what the captured answer exposed. It does not show that ChatGPT Search used no source.
More broadly, visible citations show what an answer chose to display. They do not show what the system retrieved or why it wrote what it did.
Descriptive results by engine
The study was not designed to rank engines. Each cell below is one run per product on one date, and the results do not measure how stable any engine’s answers are. Engines are listed alphabetically.
| Engine | Accurate | Partly accurate | Inaccurate | Total |
|---|---|---|---|---|
| ChatGPT Search | 28 | 1 | 1 | 30 |
| Google AI Mode | 6 | 21 | 3 | 30 |
| Microsoft Copilot | 7 | 10 | 13 | 30 |
| Perplexity | 18 | 11 | 1 | 30 |
| Engine | Yes | Partly | No | Applicable answers |
|---|---|---|---|---|
| ChatGPT Search | 19 | 0 | 1 | 20 |
| Google AI Mode | 15 | 3 | 2 | 20 |
| Microsoft Copilot | 9 | 0 | 11 | 20 |
| Perplexity | 19 | 1 | 0 | 20 |
One subset stood out on the collection date. All 10 Microsoft Copilot answers about UNDISCLOSED products were coded inaccurate overall. In many of them, external company-size bands or benchmark ranges were presented as the vendor’s named “current plan tiers.” This describes those 10 answers on 24 September 2026, not Copilot’s behaviour in general.
What SaaS teams should do
These steps follow from the error patterns above. The study did not test them. It cannot show that changing a page will change an AI answer. They are worth doing because buyers reading your own pages face the same gaps. Recommendation
- If the final price is private, publish the structure around it. The final number can stay with sales. Plan names, what each plan includes and where the custom tier begins do not need to.
- Keep commercial facts explicit and current. State plan names, billing units, commitments, minimums and allowances in plain text. Say whether a price is monthly, annual or annual-equivalent, and whether it is per user or flat.
- Separate direct pricing from marketplace configurations and outside benchmarks. If you sell fixed configurations through a cloud marketplace, say how they relate to direct pricing. The Vanta, Drata and Highspot cases show how the two routes can support different statements.
- Monitor for the patterns in this study. Check answers about your product for retired tiers, benchmark ranges presented as list prices, and separate products merged into one ladder. Our guide to diagnosing and correcting wrong AI product information sets out the workflow once you find one.
- Correct your designated first-party reference. Buyers need one accurate source. Fixing it is justified on those grounds alone. This study did not show that any particular page controls what an AI answer says.
Structured data is not a shortcut here. This study did not examine markup, and adding it to a pricing page does not guarantee a correct answer. What pricing schema establishes, and what it doesn’t covers that boundary in detail.
Limitations
- Purposive, balanced sample. The 30 products were chosen on purpose, with 10 in each pricing stratum. The results describe these 120 answers, not the SaaS market.
- One run per product on each engine. The study does not measure how stable answers are. A second run could differ.
- Tied to its conditions. Results reflect the collection date, the prompt, the logged-in accounts, the English locale, the Serbia location and the modes each surface displayed.
- Public evidence only. Private quotes and negotiated discounts were out of scope. A private quote can legitimately differ from anything public, and no private quote was judged.
- Sources change. Public pricing pages and marketplace listings can change after the collection date.
- Visible citations only. The study saw what each answer displayed, not hidden retrieval or model reasoning.
- Field-level “unverifiable” is not an error. Some details, such as a billing basis or allowance, were coded unverifiable when public evidence neither confirmed nor contradicted them. Those codes were not counted as errors.
- Marketplace ambiguity. Three original UNDISCLOSED products had vendor-sold marketplace pricing for limited configurations. The sensitivity check addresses this. Only listings that answers surfaced were checked; the other seven UNDISCLOSED vendors were not audited across every marketplace.
- Observational design. The study cannot establish why any answer was produced, or whether pricing disclosure caused any result.
Method and source notes
Prompt. Every run used this prompt, with only the product and vendor placeholders changed:
I’m evaluating [PRODUCT] from [VENDOR] for my company. What does it cost? List the current plans and prices, explain the billing basis and important usage or seat limits, and cite the sources supporting your answer.
Collection. All 120 runs were collected on 24 September 2026, one per product–engine pair, in fresh conversations with no follow-up and no regeneration. Sessions were logged in, with web search enabled, English locale, located in Serbia, time zone Europe/Belgrade. Displayed modes were recorded as: ChatGPT Search, GPT 6 Sol Light; Google AI Mode, Auto; Microsoft Copilot, Smart; Perplexity, Auto. All 120 runs were valid, with no void runs and no no-response outcomes.
Ground truth. Each answer was compared with the vendor’s public official pricing pages, plan comparisons and billing documentation, with dated amendments recorded where coding revealed additional evidence, such as the three marketplace listings.
Coding. Two coding passes were completed independently. They disagreed on 32 overall classifications, and each disagreement was settled against the captured answer and the evidence. No inter-coder reliability statistic was calculated. Answers were coded for overall result, absence preserved, official price accuracy, plan names, billing basis, minimums and allowances, third-party figures and their attribution, unsupported and outdated prices, omitted qualifications, visible citations and citation support.
Sensitivity check. The main measure was recalculated without the 12 answers about Vanta, Drata and Highspot, giving 53 of 68 strict passes.
Related work. This is an observational study of live answers. It differs from our earlier controlled test of which page layer an AI answer reads a price from, which used purpose-built pages. For setting up your own tracking of citations and price accuracy, see how to measure SaaS pricing schema, AI citations and price accuracy.
If AI answers are misrepresenting your SaaS pricing, plans or product limits, CoreAEX can trace the observable source layers, identify the first broken fact and build the correction and monitoring workflow.
Sources and data. Every figure on this page comes from the CoreAEX Contact-Sales Pricing Study: 120 answers about 30 purposively selected B2B SaaS products, collected on 24 September 2026 across ChatGPT Search, Google AI Mode, Microsoft Copilot and Perplexity, and coded against vendors’ public pricing evidence as of that date. Main-measure percentages use 80 UNDISCLOSED and PARTIAL answers, or 68 in the sensitivity check; citation-support percentages use 87 applicable answers; the unsupported-price applicable rate uses 109; all other percentages use 120. Final codes come from the adjudicated dataset after two independent coding passes. Private quotes and negotiated prices were outside scope.
About the author
Zarko Zivkovic is the founder of CoreAEX, building technical SEO, AEO, and AI-visibility systems for B2B SaaS companies. Connect on LinkedIn.