When an AI answer gets your firm wrong, treat it as an evidence problem, not a marketing problem. Save the answer, verify the fact, find the source, correct what you control, and recheck in a fresh conversation. A better answer in one chat does not prove other people now see it.
We laid out the full process in our article in The National Law Review. This post is the working version. It covers the log to keep, the access test to run, and what we found when we ran that test on our own site.
The short answer: keep a correction log, classify each error by where it came from, and test whether AI crawlers can load your key pages. Fix your own pages first. Ask third parties for corrections with evidence. Then recheck without supplying the answer.
Visibility and accuracy are two different jobs
GEO (Generative Engine Optimization) work usually chases one number: how often AI names your firm. Accuracy is a separate job. An answer can recommend your firm and still send the client to an office you closed last year.
That answer is a win on a visibility report. It is still wrong. Track the two separately, or a rising mention count will hide a misdirected client.
Four ways an AI answer goes wrong
Before you fix anything, decide what kind of error you have. Each type has a different fix and a different owner.
| Error type | What happened | How you can tell | Who fixes it |
|---|---|---|---|
| Source error | A page the AI read is wrong | The cited page contains the false statement | Whoever controls that page: you, a directory or a publisher |
| Answer error | The AI misread a correct source | The cited page does not support the statement | Report it to the platform, with evidence |
| Access error | The correct page exists, but the AI’s crawler can’t load it | Your page looks fine in a browser but returns an error to the crawler | Your web host or web team |
| Unresolved | The evidence doesn’t show the origin | No source displayed, or sources conflict | Keep logging; recheck later |
These are investigation categories. They describe what you can observe, not what happened inside the model.
See exactly where your profile stands right now.
Our GBP audit shows your current rank position across your market, how your profile completeness scores against competitors, and the specific gaps holding you back from the Map Pack.
Wrong and unflattering are not the same
A critical review that is accurate is not misinformation. Ask editors to correct false statements, not to remove legitimate criticism.
The same goes for history. A 2022 announcement that an attorney joined your firm was true in 2022. If an AI presents it as current, the fix is clearer dating on that page, not deleting it.
Keep a correction log
A correction log turns “our AI reputation looks wrong” into work someone can finish. Keep one row per disputed statement.
| Field | What to record |
|---|---|
| ID and date found | One row per statement, dated |
| Platform and mode | ChatGPT, Gemini, Perplexity, Copilot or AI Overviews, plus any search or model mode shown |
| Exact question | Word for word, with the location context |
| Disputed statement | Quote the wrong sentence exactly |
| Correct fact and evidence | The approved record, a regulator lookup or a current page |
| Error type | Source, answer, access or unresolved |
| Owner | One named person |
| Action taken | Page edited, request sent, report filed |
| Status | Requested, source updated, search result refreshed, answer rechecked |
| Next check date | When to run the same question again |
Keep the four statuses separate. A request sent is not an edit accepted. An edited page is not a refreshed search result. A refreshed result is not a corrected answer.
Escalate anything that touches licensing, credentials or another firm’s conduct to the attorney who owns risk management. Those are not marketing edits.
Run the crawler test
Access errors are the easiest to miss, because the page looks fine to a person. A robots.txt file that allows AI crawlers is not proof they can reach you. Hosting security rules, CDNs and firewalls can block a crawler before robots.txt ever matters.
The test takes about five minutes:
- Pick your key pages: home, attorney bios, practice areas and contact.
- Request each page while identifying as each AI crawler. Your web team can do this from a command line.
- Compare status codes. Every crawler you want should get a 200. A 400, 403 or 429 for one crawler but not for Googlebot is an access error.
- Fix the rule at the host or firewall, then rerun the test.
For your web team, one request looks like this:
curl -s -o /dev/null -w "%{http_code}\n" \
-A "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; GPTBot/1.2; +https://openai.com/gptbot)" \
Home
Two limits apply. This test checks rules based on the user agent only. Some firewalls also check the crawler’s IP address, so confirm in your server logs that real crawler requests succeed. And OpenAI runs separate crawlers for search, model training and user-triggered visits, so test each one.
What we found when we tested our own site
In August 2026, ChatGPT described 12AM Agency from fragments when asked to build a summary of us. So we ran the crawler test on 12amagency.com.
Our robots.txt allowed every AI crawler. Our host did not. A default hosting security rule returned HTTP 400 to GPTBot, CCBot, Meta-ExternalAgent, Bytespider and Omgilibot. Googlebot, Bingbot, ClaudeBot, PerplexityBot, OAI-SearchBot and ChatGPT-User all received HTTP 200. The pattern matched a stock “bad bot” list, not a policy anyone at 12AM chose.
We removed the GPTBot block on September 30, 2026. On October 19, 2026, we reran the test on three pages with twelve crawler user agents.
| Result on October 19, 2026 | Crawlers |
|---|---|
| HTTP 200 on all three pages | Googlebot, Bingbot, GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot, Applebot, CCBot, Meta-ExternalAgent, Bytespider |
| HTTP 400 on all three pages | Omgilibot |
Eleven of twelve now load. Omgilibot, the crawler for data company Webz.io, is still blocked. It is not an assistant a prospect uses, so it ranks low on our list. It stays in our correction log as an open access error until we decide on it deliberately.
What this proves, and what it does not: it proves GPTBot could not load our site before September 30. It does not prove that block caused any specific ChatGPT answer. OpenAI’s search crawler could reach us the whole time. That is the kind of line a correction log forces you to draw.
Check that the correction took
Verify the source and the answer separately.
- Confirm the corrected page is live and loads for each crawler.
- For pages you manage, an authorized Search Console user can request recrawling. Google says crawling can take days to weeks. That is not a timetable for AI answers.
- Ask the original question again in a fresh chat with memory off. Don’t mention the error or supply the right answer.
- Ask two or three related questions too. Office locations and the current legal team can come back different.
- Record what appeared, on which platform, on which date. If the AI stops naming your firm, log an absence, not a fix.
Give the job one owner
Assign one coordinator. Attorneys approve professional facts. Marketing makes the changes.
Keep an approved record of current attorneys, offices, practice areas, admissions and credentials. Review public information every time an attorney leaves, an office moves, the firm merges or a service changes. Ask your marketing partner for the correction log every month, alongside the visibility report.
Frequently asked questions
Can I make ChatGPT correct wrong information about my law firm?
Not directly. You can correct the pages ChatGPT reads, ask third parties to fix their pages, and send feedback through the platform with a reliable public source. Then recheck in a fresh chat. No agency can force a correction, and none should promise one.
Why does AI show my old office address or a former attorney?
This is the work we do for you. Every week, without exception.
Managing GBP at this level takes 6–8 hours a week when done right. Nova handles the entire system — posts, photos, reviews, Q&A, citations, heatmap tracking — so you can focus on running your business.
Usually because a page it trusts still says so. Look for an outdated directory listing, an old bio, a stale Google Business Profile or an undated announcement. Less often, the correct page exists but the AI’s crawler can’t load it.
How long does it take AI to reflect a website correction?
There is no fixed timetable. Google says recrawling can take days to weeks, and answers built from a model’s trained knowledge can lag further. Track source updated and answer rechecked as separate steps, with dates.
How do I know if AI crawlers can access my law firm’s website?
Request your key pages while identifying as each AI crawler, and compare the status codes with Googlebot’s. Anything other than a 200 for a crawler you want is an access error. Confirm in your server logs, because some firewalls also check IP addresses.
Should I ask a publisher to delete an old article that AI is misreading?
Usually not. If the article was accurate when published, ask for clearer dating or context. Ask for a correction only when the article itself is wrong.
The bottom line
Being recommended only helps if the recommendation describes who your firm actually is. Log every disputed statement, find where it came from, fix what you control, and test that AI crawlers can read the fix.
Want to see how AI describes your firm today, on questions that don’t already name you? Start with a free AI Visibility Scan. For how we measure, read how we measure AI visibility, and what we will not claim.
Measuring visibility is the other half of this work. Our companion post covers it: Being Cited by AI Isn’t Being Hired.
This post expands on What to Do When AI Gets Your Law Firm’s Information Wrong. Robert Portillo wrote it for The National Law Review, which published it on October 19, 2026. Our earlier NLR article covers measurement: Being Cited By AI Is Not the Same as Being Hired.



