Everybody is confusing the AI.
Nobody is doing it on purpose.
Your buyers ask AI about you before they ever visit you. The AI answers from everything you've ever published, and what you've published contradicts itself: two prices for the same plan, three customer counts, an uptime promise your own legal terms decline. We printed the receipts. Watch:
We read 61 B2B websites the way the AI does. Then we printed what we found.
Real companies, from 60-person startups to 20,000-person public companies. We kept only what we could prove: two exact quotes, from two live pages, matched character for character, with verification dates.
All 61 of 61 contradicted themselves on checkable facts. 133 verified contradictions, each one checked by a person.
Selected for B2B relevance, not randomly sampled. Companies anonymized throughout, deliberately: this is a structural problem, not a wall of shame. Full method and limits at the bottom.
The evidence file
The eight ways websites drift
- Same stat, different numbers.49
One fact, two live values on two pages. - Pricing page vs. its own docs.32
The help center sells a different deal. - One page contradicting itself.17
Hero vs. stat tile. One URL, two truths. - Renamed product, old name still for sale.8
A live page still sells the previous era. - Deadlines that already passed.6
"By 2025." Still live, still future tense. - Marketing vs. legal.5
The pricing page promises what the terms decline. - Metadata vs. visible text.5
Machines told one thing, humans shown another. - Orphaned assets.2
PDFs nobody owns, still quoting retired pricing.
133 findings total, including 9 that fit none of the eight patterns cleanly.
The facts move while you watch
On July 30 we documented a help center selling a seat for $12 that the same company's pricing page sold for $16. Five days later, the help-center price table had been replaced with an image, unreadable to machines, while a $15 worked example elsewhere still contradicts the $16 page. The fact we cited drifted before our follow-up email went out. A website is not a document. It is a population of claims moving at different speeds, and nobody is watching all of them.
Check your own site in ten minutes
The self-audit
You don't need us for the first pass. Three checks, in order of effort:
Ask ChatGPT: "What does [your company] cost, and what's included?" Compare its answer, and its sources, to what you'd tell a buyer.This is the exact experience your buyers get today.
Take your flagship stat and search your own site for it. Count how many different values are live.The study average was more than two.
Open the oldest help article and the oldest PDF Google still indexes for your domain. Read them as a buyer would.This is the step that doesn't scale by hand. It's also where the worst findings in this study came from.
If all three come back clean, you're the first. Genuinely, tell us. We'll buy you coffee and update the record.
This study was compiled by Salespeak's GTM Context Layer: a system that does the third check continuously, keeps one verified version of every fact you publish, and flags drift the day it opens. For your team, and for the machines answering your buyers.
Get your site's drift reportNo form. A real scan of your live pages, back in about a day.
- Method: every finding is two exact quotes from two live fetches, matched verbatim, with dates. Nothing paraphrased, nothing secondhand. Re-verification dates shown where a claim was checked more than once.
- Limits: 61 companies selected for B2B relevance, not randomly sampled; we make no population claim. Finding counts are floors: they scale with how hard we looked (quick passes averaged ~1 finding, deep scans ~5).
- Anonymity: deliberate, and load-bearing. The point is the pattern, not the companies. One company came within a single finding of clean; that is the record.