A rising visibility score can make a generative engine optimization program look like a win. It can also hide a shifted denominator, a confident description of a service you don’t sell, or the same lead counted twice.
Ask for the evidence behind the headline
Pick one headline claim and ask the author to trace it to saved answers, counting rules, and a calculation you can reproduce. In a 2025 Tow Center study, researchers ran 1,600 queries across eight AI search tools; the tools answered incorrectly more than 60% of the time on a source-identification task. That covers one specific task, not all AI search, but a source URL in an answer still doesn’t prove the answer represents the page accurately.
Request a test log with prompts, dates, products, modes, completed responses, failures, and retries. If the original evidence is unavailable, mark the claim unverified.
Recalculate the percentage
A denominator change can improve performance without adding a single favorable answer. If 18 of 90 eligible answers name your company, the rate is 20%. Remove 30 answers that name no brands, and 18 divided by 60 becomes 30%. That second number measures presence among brand-naming answers only; it cannot be presented as the first metric.
Keep failures visible. Check retry rules for selection bias, and count each answer once rather than counting repeat prompts as separate questions.
Check accuracy separately from confidence
NIST defines confabulation as “the production of confidently stated but erroneous or false content.” Build a dated fact record from approved service descriptions, product docs, pricing, and location details. Mark each sampled claim as supported, contradicted, unsupported, outdated, or ambiguous. Also state whether your accuracy score counts claim errors or answers containing errors.
Reconcile lead claims
If 12 unique inquiries have tracked AI referrals and eight customers mention AI during intake, with five in both groups, the combined record is 12 + 8 − 5 = 15 unique inquiries, not 20. Google Analytics documentation notes that attribution models can divide credit for one event across interactions, so channel credit and unique customer counts need different math.
Match the conclusion to the evidence
“Recorded appearances increased” and “our work generated additional revenue” require different proof. A 2019 study of 15 Facebook advertising experiments found observational estimates often failed to recover randomized effects across 500 million user-experiment observations. The lesson is methodological: detailed records alone do not establish causation. If a page update is credited with visibility gains, check timing, model updates, and competitor activity.
Run the checklist
- Can you reproduce the calculation from saved records?
- Are failures, exclusions, and retries disclosed?
- Can you verify claims against dated evidence?
- Have disputed judgments been reviewed?
- Are customer totals reconciled across sources?
- Does the conclusion stay within what the evidence establishes?
If a headline cannot be verified, keep it out of budget and performance decisions until the supporting records arrive.
Source: Influencer Marketing Hub
