Skip to content
Skip to content

What should we check when an AI answer gets a fact wrong?

Broadcastwell measures and fixes whether AI search engines name B2B software companies.

Run the free check

If you have an AI answer with a disputed fact, check the exact statement against the vendor's dated evidence before deciding what needs correction. Wrong About You classified 2 of Procore's 57 extracted statements as wrong: the underlying answers were captured on 22 September 2026 and the vendor's pages read on 8 October. Both flagged records still show team review in progress; they are evidence to inspect, not a claim that a correction has already been completed.

Observed answers on stated dates, not a ranking.

What we measured

Measured. The Wrong About You classification method records 57 Procore statements: 38 true, two wrong, none stale and 17 unverifiable. These are extracted statements, not the 109 answers that named Procore. The two wrong statements each appeared in one of three runs of the small-general-contractor comparison question.

One Claude record asserted that Procore had no pricing page. The saved vendor page contradicts that assertion. One Google AI Overview record described per-user pricing; the saved Procore source says adding users does not incur a charge. Each flag retains its question, engine, run, answer date, vendor proof and actual review state.

The classifications do not prove that either answer is still being shown or that a proposed repair will change it.

One retained answer

In one retained answer record, Google AI Overviews responded to “Procore vs Buildertrend, which is better for a small GC” on 22 September 2026, run 2, with the fragment “Per-user/percentage of revenue”. This historical wording is an answer observation, not independent verification of the product claim or an explanation of the aggregate count.

What this means for you

Reasoned. The Claude flag concerns whether a pricing page exists, not whether it publishes a fixed rate card. A correction should preserve that distinction without inventing a fixed price.

For the Overview flag, the vendor's adding-users passage addresses the per-user assertion. It does not settle every component of a Procore contract.

Procore's 17 unverifiable statements remain unresolved, not wrong. Its zero stale classifications cannot establish that no product fact has changed since capture.

The Claude pricing-page flag records Projul's Procore-versus-Buildertrend comparison as a cited page carrying the disputed statement. That gives the reviewer a named page to inspect alongside Procore's saved pricing page. The study's causing-page label does not prove that Claude read the Projul page as the cause of its answer.

An agency handover needs the in-progress team-review status. Classification, correction review and a changed served answer are separate records; neither Procore flag is marked repaired.

A separate task-level accuracy record

The Agent Gap method and release records one incorrect initial Describe task among seventy on 6 October 2026. The AppFolio session ended that day with found and reached both true, while correct was false. That is a task classification. It is separate from the Procore statement classifications above; neither base can be used as the other's denominator. It does not identify a clone site as the cause of the error or measure a branded-search defense.

Reported. Profound's FactCheck description, read on 29 October 2026, describes comparing factual claims with a company-controlled Knowledge Base, displaying source URLs, and offering proposed remediation functions. These are supplier product statements, not a test we conducted. The description distinguishes factual, actionable claims containing one verifiable assertion from subjective statements routed to Sentiment.

Profound says its Knowledge Base can be built by crawling a domain, uploading files or syncing Google Drive and Notion content. It says a claim view displays the ground truth and where it is defined. Those reported functions describe artifacts a reviewer could inspect. They do not independently establish that a particular URL caused a model's answer or that a proposed correction changed it.

Reasoned. For a suspected lookalike, keep the exact disputed claim, its cited domain and URL, and the dated official company reference together. A retained URL and a contradictory reference make the claim reviewable. Access alone does not settle accuracy, as the AppFolio task illustrates, and neither that task nor a supplier's attribution feature establishes lookalike causation. The AI source standard provides a separate public reference for the supporting source.

Also asked as

What should we check when an AI answer gets a fact wrong?

How can I defend branded search results against lookalike sites that confuse LLMs?

Lookalike clones are confusing LLMs and stealing my branded SEO. How do you defend against this?

Next step

Free 10-question check (one engine). No email is needed; runs are limited per network and per day, and by the site's daily allowance.

Claiming or correcting a vendor record is free and never depends on a purchase.

Our answer-study method, v1.1.

Answer Repair: $490.

Prepared by the Broadcastwell team.

Page owner: Marcus Reed, AI Specialist. Corrections: hello@broadcastwell.com