Find What AI Gets Wrong About You

Missing from an AI answer costs you a buyer you never meet. Being in the answer with the wrong facts can cost more. The buyer reads a price you stopped charging, a feature you never built or a CEO who left two years ago, and either crosses you off or turns up on a sales call expecting something you do not sell.

Nobody tells you when this happens. The answer is generated for one person, read once and gone. The only way to find it is to collect answers and grade them against what is true.

This lesson covers the errors worth checking for, why engines make them, how to trace a wrong fact to the page it came from, how to find the contradictions on your own site, and how to decide what to fix first. You finish with a canonical fact sheet and a graded set of answers.

The errors worth checking for

Not every inaccuracy matters. Check for the ones a buyer acts on. These are the categories the 99Visibility accuracy check covers, and they make a sound manual checklist too. The examples use Ledgerly, our fictional B2B invoicing SaaS.

Error	What it looks like for Ledgerly	Why it costs you
Pricing	"Starter costs $25 per user." It has been $29 since April 2026.	Sets the wrong expectation before the first call
A feature you have, denied	"Ledgerly does not support recurring invoices."	Drops you from shortlists you belong on
A feature you lack, claimed	"Ledgerly includes payroll." It does not.	Buyers arrive expecting it and leave angry
Founding date	"Founded in 2021." It was 2019.	Makes you look newer and riskier
Location	"Headquartered in Denver." It is Austin.	Confuses tax, compliance and local questions
CEO and leadership	"CEO Tom Hale." Priya Nair has led since 2024.	Signals the engine's picture of you is stale
Ownership	"Ledgerly owns Ledgerly Time." It was sold in 2025.	Sends buyers to a product you no longer support
Acquisitions	"Ledgerly was acquired by Invoxa." It never was.	Tells buyers you are someone else's product
Headcount	"Ledgerly employs about 400 people." It has about 60.	Misleads buyers about your scale and support

Outdated facts deserve their own label. A former CEO named as current, or a product you sold described as yours, was true once. That tells you where it came from: an old page, an old article or a training snapshot, not pure invention.

The 99Visibility check compares every answer with the brand facts you enter, and it only reads claims whose subject is your brand, so a competitor's founding year in the same answer never counts against you. It reads specific claim shapes: a dollar price, a founding year, a "headquartered in" place, a named CEO, an "acquired by". An error phrased some other way can slip past it, so read the answers yourself too. Two kinds of claim get flagged as unverifiable rather than wrong: a specific headcount when you have no employee fact on file, and an acquisition your facts say nothing about. Unverifiable means your fact sheet has a gap, not that the engine is lying.

Why AI gets your facts wrong

Four causes produce nearly every wrong fact. You control one of them completely and can influence two more.

A stale snapshot

An answer from memory reflects a model snapshot. In this fictional example, Ledgerly changes its price after the measured model’s cutoff, so that snapshot cannot have learned the update from Ledgerly’s new page. Correct the sources, disclose the model and its cutoff, and separate this from a live search result. OpenAI’s chat-latest model page, checked 16 October 2026, lists an April 20, 2026 knowledge cutoff. That alias changes over time. Other models have their own cutoffs, so record the exact model and its documented cutoff rather than treating ChatGPT as one fixed knowledge state. A fact published after that snapshot cannot have been learned from that publication during training.

Contradictory sources, including your own site

If your pricing page says $29, a 2024 blog post says $25 and your structured data says $27, an engine that reads all three has no way to know which is current. It repeats one of them. This is the plausible mechanism behind most "AI gets our pricing wrong" complaints, and it is the one cause fully under your control. Nobody has published a controlled test of it, so treat it as the first thing to check, not as a proven law.

Third-party pages

Review sites, listicles, directories, old press coverage and competitors' comparison pages all describe you, and every engine that searches the web reads them. A 2024 review with your old price is a live source for every retrieval-on answer until someone updates it.

Injected content

Some pages carry hidden instructions aimed at AI systems. A 2026 study of agentic search (arXiv:2606.16821) found that manipulated content placed in a mock search setup could change answers, and that attack success varied widely by model and by setup. The study never published anything to the open web, so it shows what happens once such a page is retrieved, not how often that occurs. Treat it as a defensive risk: if answers about you are wrong in a specific, repeated, odd way, read the pages they cite.

Trace the source in retrieval-on answers

When an answer shows sources, you can often find where a wrong fact came from. Work through it in order:

1. Open every cited page and search it for the wrong value, such as "$25" or "Denver".
2. If a cited page states the wrong fact, that page is your target. If it is yours, fix it. If it belongs to someone else, ask the owner to correct it and link them to your canonical page.
3. If the cited pages state the right fact, the engine blended or misread them. Check whether your own page states the fact plainly, in text a crawler can read.
4. If no cited page mentions the fact, it came from somewhere uncited or from memory. Citation is not use: OpenAI says the number of sources consulted is often greater than the number of citations shown.

Then check reach. If your correct pricing page loads its plan table with JavaScript, many AI crawlers never see the price. Vercel's crawler study (December 2024) found the ChatGPT and Claude crawlers fetch JavaScript files but do not execute them, and independent re-testing in June 2026 found the same for PerplexityBot. A page blocked by robots.txt or by a firewall is equally unreadable. Module 3 covers each check: robots.txt bot by bot, firewalls and CDNs, and rendering and snippet controls.

Answers with no sources have nothing to trace. Record them, note the model's cutoff, and work on the sources the next model will learn from: your own pages first, then the third-party pages that show up when you search your own brand name.

Audit your own site for contradictions

Before you blame an engine, make sure your own site tells one story. Take each key fact and find every place it appears:

• The pricing page, plan comparison tables and checkout
• The homepage, about page, footer and press kit
• Help center and documentation, including old articles
• Blog posts, changelogs and launch announcements
• Structured data (JSON-LD) and meta descriptions
• PDFs, sales decks and app store listings
• Profiles you control elsewhere: LinkedIn, Crunchbase, review-site listings

Search for the old values, not the current ones. A site search for "$25", "2021" or the former CEO's name finds the pages that still say them. Your CMS search and a search engine's site: operator both work. Compare your structured data with the visible text using the free schema generator and parity check: a price that differs between the two is a contradiction you published yourself.

Resolve each contradiction the same way: one current value on every live page. Update the pages that should be current. Leave historical posts historical, but put the date in the text, such as "Prices as of March 2024," so a reader and a crawler can both tell old from current. Put the current value in server-rendered HTML. Do not bump a page's modified date to look fresh. Engines can see such dates, but nobody has shown they weigh them, and a date that changed without the content changing proves nothing.

Write the fact sheet, then triage

Grading answers from memory does not work. Two people on your team will disagree about the founding year. Write a canonical fact sheet: one value per fact, the date it took effect, and the page that states it. This is Ledgerly's.

Fact	Current value	Stated on
Starter price	$29 per user per month, since April 2026	ledgerly.example/pricing
Growth price	$49 per user per month	ledgerly.example/pricing
Trial	14-day free trial, no free plan	ledgerly.example/pricing
Founded	2019	ledgerly.example/about
Headquarters	Austin, Texas	ledgerly.example/about
CEO	Priya Nair, since 2024	ledgerly.example/about/team
Ownership	Independent, never acquired	ledgerly.example/about
Divested	Ledgerly Time, sold in 2025	ledgerly.example/newsroom
Integrations	QuickBooks and Xero	ledgerly.example/integrations
Not offered	Payroll	ledgerly.example/features

Grade every stated fact in your audit sheet as correct, wrong, outdated, unverifiable or not stated. Then enter the same facts in your 99Visibility brand profile so the accuracy check has something to compare against. Without them it checks nothing.

You will find more errors than you can fix in a week. Rank them on three questions.

How bad is it? The 99Visibility accuracy score weights errors by severity. An acquisition claim your facts contradict is critical. Wrong pricing, a denied or invented feature, a wrong location and the wrong current CEO are moderate. A wrong founding year, a wrong CEO start year and an unverifiable headcount are minor. Use the same order unless your business says otherwise.

How often? One wrong answer in one run is a lead. The same error on three engines across two runs is a pattern.

Can you reach the cause? A retrieval-on error traced to your own page is a fix for this week. One traced to someone else's page is an email. An error from memory with no source is slow: correct the sources and re-measure after the next model release.

Common mistakes

• Grading without a written fact sheet. You will mark a wrong answer correct because it matched what you half-remembered.
• Trying to correct the engine instead of the source. You cannot edit an answer. You can fix the pages engines read.
• Searching your site for the current value. Search for the old ones; those are the pages that contradict you.
• Treating every error as equal. An invented acquisition and a founding year off by two are not the same problem.
• Calling an unverifiable claim false. It is a gap in your facts. Add the fact, then re-check.
• Leaving brand facts empty in 99Visibility. Accuracy then stays at 100 and checks nothing.
• Expecting a site fix to change an answer from memory this quarter. Retrieval-on answers can follow your fix; frozen snapshots cannot.