When an AI engine describes your product incorrectly, it presents a confident, plausible, wrong answer to buyers evaluating you and it continues doing so until the underlying sources it drew from are corrected.
There is no support ticket. No dashboard flags it. No one at the company that built the model reviews your case. The correction happens indirectly, by changing what the system reads, and it takes weeks.
This is a different problem from being absent, and in most cases a more urgent one. Absence means buyers do not consider you. Misdescription means buyers actively rule you out for a reason that is not true.
How common is this?
Misdescription is common enough that it should be checked before any visibility work begins, and it is more likely for companies that have changed pricing, positioning or feature set in the last two years.
The pattern in practice: when running the direct prompt “is good for [use case]” across B2B software companies, results fall into three groups. Some are described accurately. Some produce nothing at all. And a meaningful proportion return something wrong.
Companies most at risk:
- Recently repositioned: You moved from one category to another; third-party sources still describe the old one.
- Recently changed pricing: Old figures persist in review sites, roundups and comparison articles.
- Sunset a feature: Articles describing it remain indexed.
- Rebranded or been acquired: Entity confusion between old and new names.
- Have a name similar to another company: The model conflates you.
What kinds of errors occur?
Six error types, in rough order of commercial damage:
1. Incorrect compliance or security claims: In regulated categories, a model stating you hold a certification you do not or omitting one you do is a serious problem in both directions. The first creates a false expectation that surfaces badly in procurement. The second removes you from consideration by enterprise buyers who filter on it.
2. Outdated pricing: Frequently the most damaging routine error. A buyer told you cost $200 per seat when you now cost $60 does the maths and rules you out before ever visiting your site.
3. Wrong category: You are described as a CRM when you are a sales engagement platform. This removes you from every query in your actual category and inserts you into ones where you compare badly.
4. Discontinued features described as current: Creates expectations your product cannot meet, which surfaces in demos as a credibility problem.
5. Missing capabilities you actually have: The model has not learned about a feature you launched eighteen months ago because nothing crawlable describes it clearly.
6. Inappropriate competitor comparisons: Being repeatedly compared against a company in a different segment shape how buyers frame their evaluation of you
How do you find out?
Run the direct prompt and check every factual claim, not just whether you appear
The prompts to run in fresh sessions, across ChatGPT, Perplexity and Google AI Overviews:
- Tell me about [your product]
- What does [your product] cost?
- Who is [your product] for?
- What are [your product]’s main features?
- Is [your product] good for [your primary use case]?
- How does [your product] compare to [main competitor]?
- Is [your product] SOC 2 compliant? (or your relevant certification)
How to record it: For every response, note each factual claim and mark it correct, incorrect or outdated. Screenshot everything with the date visible. You need this baseline to know whether any correction worked.
Run each three times in separate sessions: These systems produce variable output. A wrong answer once may be noise; a wrong answer three times is a pattern with a source behind it.
How do you trace the source
Use Perplexity, which displays its citations, to identify which sources the incorrect claim came from
This is the most useful diagnostic step and the one most people skip. Perplexity shows the URLs it drew on. Run the prompt that produced the error and read the cited pages.
In most cases the source is one of four things:
- Source Typical error
- An outdated third-party article Old pricing, discontinued features
- Your own stale content A page you forgot to update after repositioning
- A review platform listing Wrong category, outdated feature list
- A competitor’s comparison page Unflattering or outdated characterisation of you
The fourth is worth dwelling on. Your competitors publish pages describing you, those pages get read by these systems, and if you have no equivalent page of your own, theirs is the only version available.
If ChatGPT produces an error that Perplexity does not, the claim is likely in trained knowledge rather than live retrieval. That is harder to correct and updates only on model releases which makes fixing the underlying sources more important, not less, since those sources feed future training.
How do you correct it
Correct the sources, not the symptom. There is no direct channel to the model:
Step 1: Fix your own pages first: State the correct fact plainly, in crawlable text, with a visible date. Do not bury it in a paragraph. A model needs an extractable statement:
“As of [month year], [Product] pricing starts at $60 per user per month. Previous pricing tiers announced before [date] are no longer current.”
Explicitly naming the outdated information helps, because it gives the correction something to attach to.
Step 2: Update every third-party listing you control: G2, Capterra, Trust Radius, Crunchbase, LinkedIn, Product Hunt, industry directories, your own documentation. Most companies have eight to fifteen of these and have not touched them in two years.
Step 3: Contact publishers where a third-party article is wrong: Politely, with the correct information and a link to your source page. Success rates are moderate but not negligible most publishers will correct a factual error if you make it easy for them.
Step 4: Build the comparison page if a competitor’s version is the source: If their characterisation of you is the only one that exists, the fix is to publish your own.
Step 5: Wait, then recheck monthly: Live retrieval updates in weeks. Trained knowledge updates on model releases. There is no faster path, and anyone claiming one is selling something.
What if the error is about a competitor comparison?
If a model repeatedly compares you against companies, you do not consider competitors, the fix is to publish comparisons against the ones you do.
This is worth understanding as a positioning signal rather than an error. The model is reflecting the consensus it found. If the web consistently mentions you alongside a company in a different segment, that is what buyers are being told, and it probably reflects how third parties actually perceive you.
Two responses, and you likely need both:
Short term: Build comparison pages against your actual competitors so the correct comparisons exist in extractable form.
Longer term: Treat it as evidence that your positioning is not landing externally. If you describe yourself one way and the entire web describes you another, external consensus wins in these systems and probably in buyers’ minds too.
How long does correction take?
Live retrieval typically reflects corrections within four to eight weeks. Trained knowledge only updates when a new model version is released
A realistic sequence:
- Weeks 1–2: Your own pages corrected, third-party listings updated. Nothing changes in AI answers yet.
- Weeks 3–6: Corrected pages crawled and incorporated. Perplexity and other retrieval-heavy engines begin reflecting the change.
- Weeks 6–12: Broader improvement as third-party corrections publish.
- Next model release: Trained knowledge updates outside your control and unpredictable in timing.
Recheck monthly using the same prompts and record the change. Partial correction is common: an engine may state the correct pricing while still describing a discontinued feature.
Why fix this before working on visibility:
Increasing the reach of an incorrect description makes the problem worse.
If a model describes you in the wrong category at low visibility, few buyers see it. Successfully improving your visibility means more buyers see the wrong description, and they rule you out with more confidence because it came from a source they trust.
The sequence that makes sense:
1. Check accuracy
2. Correct any errors and confirm the correction has landed
3. Then work on visibility
This is a short delay four to eight weeks and it prevents amplifying a problem.
Frequently Asked Questions:
Can I contact OpenAI or Google to fix incorrect information?
There is no reliable direct correction channel for factual claims about a company. Some platforms have feedback mechanisms, but the practical route is correcting the underlying sources.
How do I know which source caused the error?
Perplexity displays its citations. Run the prompt that produced the error and read the cited pages. In most cases the source is identifiable within a few minutes.
What if my competitor’s page is the source of a wrong description?
Publish your own comparison page with accurate information. You cannot compel them to change theirs, but you can ensure a correct version exists in extractable form.
How often should I check for this?
Monthly, as part of the same prompt testing you run for visibility. Add the accuracy check to the existing routine rather than treating it separately.
Does correcting my own site fix the trained knowledge
Not immediately. It fixes live retrieval within weeks and improves the sources that feed future training, but trained knowledge updates on model releases.
Frequently. Absence means buyers do not consider you. Misdescription means they actively rule you out based on something untrue, and they do so with confidence.