AI Product Claims Study 2026: Evidence From 12 Launches

Virgo PR reviewed 12 AI product announcements from September 2026 and found that all 12 explained a concrete capability, 10 stated an access or release state, and 4 named an outside user or evaluator. Three announcements made five numerical product claims, and none of the five fully disclosed how the number was measured. This is a small convenience sample, not a ranking or a market estimate.

Updated November 10, 2026. Announcements dated September 2 to 30, 2026. Pages accessed November 10, 2026.

What did the study find?

An AI product announcement can tell readers what a system does without giving them enough information to judge how well it performs. Virgo PR examined that gap in 12 announcements from 12 companies across enterprise software, marketing technology, cybersecurity, financial technology, healthcare technology and consumer AI. The five principal findings are below.

• Capability was clear in 12 of 12 announcements (100.0%). Each named a practical task and a user or use case.

• Access or release state was identifiable in 10 of 12 (83.3%). A clearly stated planned release counted as clear disclosure, but it does not make a product generally available.

• Named outside use or evaluation appeared in 4 of 12 announcements (33.3%). After a bounded supporting review, that rose to 5 of 12 (41.7%), and the added record was a later hands-on evaluation, not production adoption.

• Numerical product claims appeared in 3 of 12 announcements (25.0%), contributing five separate claims. Two of the three comparative claims named their baseline fully, and one did so partially.

• Relevant independent evaluation was found for 1 of 12 companies (8.3%) through supporting search, after launch and not in the announcement itself.

How do announcement-only and supporting evidence compare?

The table separates what each announcement said on its own from what a bounded supporting review added. The denominator is 12 announcements in every row.

Announcement-level dimensionAnnouncement onlyAnnouncement plus supporting review

Clear capability and intended use12/12 (100.0%)12/12 (100.0%) Clear access or release state10/12 (83.3%)10/12 (83.3%) Named outside use or evaluation4/12 (33.3%)5/12 (41.7%) Qualifying numerical product claim3/12 (25.0%)3/12 (25.0%) Relevant independent evaluation0/12 (0.0%)1/12 (8.3%)

The claim inventory was fixed to the announcement, so supporting pages could substantiate a numerical claim but could not add new ones.

How well did the numerical claims disclose their basis?

The five numerical claims are the denominator for method and conditions. Baseline applies only to comparative claims, and two of the five were absolute figures with no comparison to disclose.

Claim-level disclosureFullPartialNot found in defined searchNot applicable

Baseline (comparative claims)2/31/30/32/5 absolute claims Measurement method0/54/51/50/5 Relevant testing conditions0/55/50/50/5

The difference between full and partial matters. A workload count can make an anecdote more informative without supplying a reproducible test. Naming a metric does not explain how records were sampled, how time was measured, or whether a comparison reflects an observed counterfactual.

How was the study designed?

The completed sample is 12 announcements from 12 companies, two per sector, reduced from a target of 40 so that every record received the same primary-source inspection and bounded supporting review. It is a purposive convenience sample. It was not randomly selected, and the reduction does not imply that only 12 eligible announcements exist.

• Window and dates. All included announcements were dated September 2 to 30, 2026. Pages were accessed November 10, 2026, so this is a review of publicly accessible pages at that time and not an archived reconstruction of launch-day disclosures.

• Supporting review rule. For each company the study inspected the announcement, up to two directly linked substantive pages, exactly two targeted searches, and at most three relevant pages from those searches.

• Claim selection. The register holds 17 focal records: one primary capability claim per announcement and five distinct numerical product-result claims. Corporate scale, funding, prices, interoperability counts and market surveys were excluded from the numerical-benefit denominator.

• Coding. One researcher performed selection, coding and a second verification pass. Calculations were recomputed and featured examples rechecked against sources. There was no independent human coding panel or inter-rater reliability assessment.

Which sectors were included?

Two records per sector cannot support a league table, so this is a description of what was encountered and not a sector ranking.

SectorCompaniesNamed use in announcementNumerical product claims

Enterprise softwareDataiku; Phrase1/21/2 Marketing technologyBraze; Apollo1/21/2 CybersecurityZscaler; Proofpoint1/20/2 Financial technologyForm3; Sapiens1/20/2 Healthcare technologyOracle; Sunny Health AI0/20/2 Consumer AIInstacart; OpenAI0/21/2

What do the documented examples show?

Phrase separates workload detail from a reproducible test

Phrase's Atlas announcement describes three anonymous enterprise examples: settings reviewed across 59 templates, 66 language gaps filled, and 782 jobs checked across five languages. These are useful workload details, but they do not supply elapsed times, repeated-trial counts or a full test procedure. The named Bitly feedback is separate and should not be treated as the source of those anonymous results.

Braze shows why methodology must match the claim

Braze's announcement describes beta access, named Sportsbet use of Agentic Standards, and a capacity claim above one billion simultaneous combinations for Decisioning Studio Go. Its measurement guide explains random-control and business-as-usual comparisons. That is useful guidance for measuring engagement lift, but it does not validate the capacity claim. The page shows an October 15 edit, so its current wording cannot all be assigned to the launch date without a historical capture.

Form3 makes an operational boundary explicit

Form3's announcement describes authorized access to payment information through an MCP adaptor and specifies read-only operation, meaning the AI does not initiate changes or enter the payment flow. The bank and fintech names relate to the underlying platform and were not counted as deployments of the new AI layer.

Sapiens names evaluation without proving outcomes

Sapiens names Continental General as evaluating its AI tools against real workflows. That qualifies as named evaluation evidence. It does not establish broad production deployment or measured return on investment, and the announcement also describes human review of proposed mappings.

Instacart gives a baseline without establishing causation

Instacart reports that average Clementine orders exceed its $115 overall basket average. The review did not find the Clementine mean, cohort count, observation window or matched comparison, so the $115 figure is a baseline and not an observed uplift. Named Cart Assistant grocers were not counted as Clementine deployments.

What can communications teams change?

The practical opportunity is to attach evidence to the specific claim that needs it. More links are not automatically better disclosure. A benchmark about another task, a survey about market demand or a case study about an earlier product can be credible while leaving the launch claim unanswered. Teams can work through five habits before a release goes out.

• Separate function, access and result. Say what the product does, who can get it now or later, and what was actually observed.

• Put the baseline next to the number. Name whether the comparison is a prior workflow, an earlier version, a competitor or a control group, and state units and the measurement window.

• Keep workload and conditions near the figure. Task counts, document length, model version, integrations and human review all help a reader tell an anecdote from a repeated test.

• Attribute customer evidence to the right product. A customer using an older platform is not necessarily using the new capability, and a pilot should be described as a pilot.

• Date the evidence. Record publication and update dates, and label later evaluations as later additions so journalists do not assign them to launch day.

Virgo PR helps technology teams make these distinctions clear in their communications, from technology PR programs to launch materials that journalists can evaluate. This study does not show that Virgo clients disclose more evidence or perform better than other companies.

What are the limitations?

The study is small, English-language, search-discovered and balanced by sector on purpose, so no statistical inference to the AI market is justified. The fixed search allowance can miss evidence, and inaccessible pages and unknown publication histories remain limitations. No company was contacted, no product was independently run, and no financial-independence audit of outside reviewers was performed. Missing public evidence means not found within the defined search. It is not a finding of falsehood, deception, noncompliance, poor quality or dissatisfied customers, and no composite score or company ranking was produced.

Disclosure

Virgo PR is a public relations agency that works with technology companies and published this study. None of the 12 companies was publicly identified as a Virgo client in the two reviewed agency pages (technology practice and our work) or in company-specific site searches, and those sources are not an exhaustive client list.

Frequently asked questions

How many AI product announcements did the study cover?

It covered 12 announcements from 12 companies, all dated September 2 to 30, 2026. The original target was 40, reduced so every record could receive the same bounded review.

Does the study say any company's claims are false?

No. The study measures what was publicly disclosed within a defined review. Missing evidence means it was not found in that search, not that it does not exist or that a claim is wrong.

How many numerical claims fully disclosed a method?

None of the five. Four had partial method descriptions and one had no qualifying method found.

Can these results be applied to the whole AI market?

No. The sample is a purposive convenience sample of two announcements per sector, so it does not estimate market-wide or sector-wide rates.

What evidence counted as independent?

Only identifiable evaluation or data produced by a party independent of the company. Customer quotations, syndicated releases and company citations to market context did not qualify automatically.