How to check what ChatGPT, Gemini and Perplexity say about your company, and how to correct it
Ask the four major assistants a fixed set of questions about your company and your category, record the answers, ask each one for its sources, then fix the wrong claims where the assistants actually read them. One hour to run the first audit; a repeatable monthly check thereafter.
Why audit before optimising
Every visibility project I have seen go wrong went wrong the same way: it started with tactics instead of a baseline. An hour of structured asking tells you which of the three failure states you are in (unknown, described wrongly, or known but never recommended), and each state has a different fix. I ran exactly this exercise on my own company in August 2026, and I will use its results as the worked example below.
Step 1: build your query panel
Write down 10 to 15 questions in three groups, phrased the way a real buyer would ask. Save them; the value compounds only if you ask the same questions every month.
- Identity queries. "What is [company]?" "Who founded [company]?" "Where is [company] based and what does it do?"
- Category queries. "Recommend a [what you sell] provider for a [typical customer] in [city/India]." "Best [category] companies in [city]." These are the queries where recommendations are won and lost.
- Comparison queries. "Is [company] good?" "Compare [company] with [competitor]." "Alternatives to [competitor]."
Step 2: ask, and record everything
Run the panel through ChatGPT, Gemini, Perplexity and Claude. Use a spreadsheet with one row per query per assistant, and record: the description given, whether you were named at all on category queries, which competitors were named, anything factually wrong, and the overall tone. Tedious for twenty minutes, invaluable for the next year: this sheet is your before photo.
Step 3: ask for sources
The step almost everyone skips, and the one that converts complaints into a task list. Follow up with: "What sources is that based on?" or "Cite the pages you used." Perplexity cites by default; ChatGPT, Gemini and Claude will show or summarise sources when asked, especially when they searched the web to answer. Record the URLs. Patterns appear immediately: usually a handful of the same directories, listicles and old articles feeding every assistant's view of you.
Step 4: trace each wrong claim to its source
My own baseline made this concrete. The assistants identified me correctly as a person (that took deliberate work, described in the disambiguation case study), but several described EbizIndia as a "software development and digital marketing" company: our positioning from years ago, not the current one. The sources? Old third-party profile pages and directory entries, some of which I had forgotten existed. Not our website, and not LinkedIn. The machines were faithfully reporting an outdated record that we had left lying around the internet.
Sort what you find into three buckets: wrong on pages you control (your site, your profiles), wrong on pages you can influence (directories, review platforms, partners), and wrong on pages you can neither edit nor influence (old news articles, dead blogs).
Step 5: fix at the source, in that order
- Pages you control: fix this week. Site copy, structured data, LinkedIn, Google Business Profile, every social bio. Use one identical one-line description everywhere; consistency is itself a signal.
- Pages you can influence: claim and update the directory listings, request corrections from partners, refresh stale review-platform profiles. Slower, and worth it: these are precisely the corroborating sources assistants weigh.
- Pages you cannot change: outweigh them. Publish current, dated, retrievable material so the fresh record outnumbers the stale one. Old pages lose authority fastest when newer sources contradict them consistently.
Step 6: recheck monthly and expect this timeline
Same panel, same spreadsheet, new column. Perplexity typically reflects fixes within days because it retrieves live. Google surfaces (Gemini, AI Overviews) usually follow within two to six weeks of recrawling. ChatGPT's description of your company moves slowest, often one to three months, because it leans on the accumulated off-site record. If nothing has moved in ninety days, the usual culprits are: fixes made only on your own site while the cited third-party sources still say the old thing, or AI crawlers blocked at the server without anyone noticing, which is covered in the robots.txt guide.
From here, the correction project becomes a visibility project: once the record is accurate, the same monthly panel tells you whether you are starting to win the category queries too. That is the subject of how assistants decide who to recommend.
Questions founders ask
How often should I re-run the audit?
Monthly is enough. Assistants do not re-form their view of a company daily, and monthly gives your fixes time to be recrawled. Put a fixed date in the calendar and keep the query set identical so results are comparable.
The assistant says it cannot find my company at all. Is that worse than being described wrongly?
It is more common and easier to fix. Absence usually means retrieval is failing: the site is thin, uncrawlable or new, and there is no third-party record to fall back on. Wrong descriptions require correcting existing sources; absence requires creating them.
Can I just tell ChatGPT the correct information in the chat?
It will politely accept the correction for that conversation and forget it for everyone else. Corrections only persist when made at the sources assistants retrieve from: your site, directories, review platforms and the pages that mention you.
Should I use the paid versions of the assistants for the audit?
Use whatever your buyers use, and run queries in a fresh session without personalisation where possible. The paid tiers search the web more readily, which is closer to how a serious buyer researches, so testing on both free and paid tiers is ideal.