There's a specific failure mode when AI writes your competitive analysis: it's fluent, confident, plausible, and occasionally describing pages that don't exist. A model asked "what do these competitors cover that this site doesn't?" will produce a beautiful list whether or not it actually verified anything. The prose gives you no way to tell which items are observations and which are vibes.
This matters because gap analysis drives spending. You'll commission pages, brief writers, allocate weeks based on this list. A hallucinated gap costs you a page nobody needed; a hallucinated "covered" hides a hole that keeps costing referrals. Either way you paid for fiction.
The fix: verdicts that can only point at real pages
AnvixaAI's gap runs are built so unverifiable claims can't exist, structurally. Before any AI is involved, the platform builds two catalogs of real, fetched pages: your site's crawled inventory, and a fresh light-scan of each competitor (their sitemap plus their key pages). The model's job is constrained to mapping: organizing those catalogs into customer-language topics. And when it renders a verdict, the evidence it's allowed to cite is catalog entries only: it points at pages by index, not by description.
Code then grounds every verdict back to the catalog. The result is mechanical honesty:
- A covered verdict carries the URL of your page that covers the topic. Click it; it's a real page you already have.
- A missing verdict carries the competitor pages that answer it. Click those too and see exactly what an assistant sees when it goes looking.
- A topic with no evidence on either side simply can't be asserted. The model has no mechanism to cite a page that wasn't fetched.
Every verdict resolves to a fetched, clickable page. No receipt, no claim: the mechanism, not a policy.
Receipts change how you use the report
Disagreement becomes productive. Think a "missing" is wrong because your blog post covers it? Open the receipt: maybe the topic is genuinely buried in paragraph nine of an unrelated post, which is its own finding, because if the analysis couldn't match it, neither can an assistant answering a customer at speed. What reads to you like a false negative is usually a discoverability problem wearing a disguise.
Briefs write themselves. A missing topic ships with the competitor pages that answer it today. Hand both to whoever writes the page: here's the question, here's the bar, clear it in our voice with our expertise. No research phase.
Trust survives repetition. Run the analysis again next quarter and verdicts move for checkable reasons: you published a page, a competitor added a section. An unverifiable report you read once and shelve; a receipted one becomes an instrument you re-read.
- Fluent AI analysis can describe pages that don't exist, and you can't tell which
- Constrain the model to mapping real, fetched pages; let code ground every verdict
- Covered → your page's URL; missing → the competitor pages that prove it
- Clickable evidence turns disagreements into findings and reports into instruments