AI answers say things about your company that you never wrote. Some of them are wrong: an old price, a feature you shipped months ago listed as missing, a product you never made. A buyer who reads that sentence does not check it. They move on to the next name on the list.
This post is about fixing those sentences. Not tracking them, fixing them. The short version: a wrong answer is almost always repeating a page somewhere, and the page is the thing you can change.
What counts as a misrepresentation in an AI answer?
A misrepresentation is a checkable statement about your brand that is false. "Acme starts at 49 dollars a seat" is checkable. "Acme is the easiest tool to use" is not; it is an opinion, and opinions belong to brand sentiment tracking, not here.
Most wrong claims fall into five kinds:
Pricing
- Example of a wrong claim
- A plan price from last year
- Usual source
- An old pricing page, a review site's pricing tab
Features
- Example of a wrong claim
- "Does not support X" after X shipped
- Usual source
- Your own comparison page, a stale listicle
Integrations
- Example of a wrong claim
- A connector you dropped, or one you never had
- Usual source
- Partner directories, old docs
Company facts
- Example of a wrong claim
- Wrong founding year, country or owner
- Usual source
- Business databases, press releases
Identity
- Example of a wrong claim
- Your name confused with another product
- Usual source
- Name collisions, inconsistent naming
The kind matters because it tells you where to look first. Pricing and feature errors usually come from a page that was true once. Identity errors usually come from a page that was never about you at all.
What did AI answers get wrong about Attensira?
We read our own brand questions every day in our own workspace, so we can show this on ourselves rather than on a made-up company. These are answers read on 2026-10-03, quoted from the stored answer text.
ChatGPT, "Attensira vs Profound", three samples. Two of the three samples put a table in front of the reader with a row for crawler analytics, marked "No" for us. That was wrong on the day it was written: AI crawler traffic is a shipped feature, on Growth and above, documented on our crawler logs page. One sample also listed two engines among the ones we read that are still marked as coming, not live.
Where it came from. Every one of those samples cited the same page first: our own old comparison page. The answer even said so: "Attensira's own comparison says Profound has broader engine coverage and stronger crawler/agent analytics." The wrong claim was ours. We had written a comparison before crawler analytics shipped and never went back.
Perplexity, "what is Attensira", three samples. Two of the three samples opened by saying the name can refer to two different things, and described an ADHD medication with a similar name next to the software. That is an identity error, and the fix for it is different. We wrote it up in entity fog.
Perplexity, "Attensira vs Profound", three samples. Two of the three samples declined to compare at all, saying the pages it found described us "in vague terms". That is not a false claim. It is a missing one, and it has its own fix: a page that states the facts plainly enough to be quoted.
Three questions, nine answers, three different problems. That is the normal shape. "AI is wrong about us" is never one problem.
Where does a wrong AI answer come from?
An answer engine writes from two things: pages it reads while answering, and what the model learned before. You can see which one it used by looking at the citations.
The answer cites a page for the wrong sentence. This is the easy case and the common one. The page says the wrong thing, or says something old that the engine reads as current. Fix the page.
The answer cites pages that do not contain the claim. The model blended several sources, or filled a gap from memory. Read every cited page for the fact. Often one of them says something close enough to be misread, such as a table with a blank cell or a "coming soon" line with no date.
The answer cites nothing. The claim came from what the model learned earlier. You cannot edit that directly. What you can do is make the correct fact consistent and easy to find everywhere the model will read next time: your site, your docs, your listings, your profiles.
The answer confuses you with someone else. Two entities share a name or something like it. The fix is naming and disambiguation, not a correction of a fact.
How do you find every wrong claim, not just the one you saw?
Spotting one wrong answer by chance tells you nothing about how often it happens. Do it on purpose.
- Write down the facts that cost you a sale when they are wrong. Price, plans, what you integrate with, what you do not do, where you sell, who you are. Ten to twenty facts is enough.
- Ask the questions buyers ask, on every engine you care about. "What is [brand]", "[brand] pricing", "[brand] vs [competitor]", "does [brand] support [X]". Ask each more than once; answers vary from sample to sample.
- Copy the exact sentence for every claim about you. Not a paraphrase. The exact words are what you will search for on the cited pages.
- Check each sentence against your list of facts. Mark it right, wrong, or nothing to check against.
- Repeat on a schedule. A claim that was right last month can go wrong when a review site updates its table.
The manual version works for a dozen questions. Past that it becomes a job. In Attensira this is the Claims about you check: it pulls each checkable statement out of every answer it reads, keeps it only if the quote appears in the answer's own text, and checks the ones about you every night against the facts you confirmed in your brand profile. A contradiction is raised only when the verdict's confidence is at least 80 percent, and a fact the agent drafted from your site is never used until a person confirms it. The claims documentation has the full rules.
