Short answer: it depends entirely on who is asking. Ask ChatGPT "is [brand] a good CRM" and you get an answer built largely from review platforms and your own documentation. Ask "is [brand] a good place to work" and your website nearly disappears from the source list, replaced by Glassdoor, Reddit and job aggregators. Same company, different retrieval path, different conclusion.
Most teams that ask what does AI say about my company only test the first kind of question — the buyer prompt set. That’s roughly a quarter of what’s being asked about you, and the only quarter that pulls heavily from pages you control.
This article covers all four audiences: what each one asks, which sources feed their answers, the citation data behind the split, and how to audit and fix your own in an afternoon.
What does AI say about my company?
AI assembles a different answer about your company for every audience. Buyer answers draw about a third of citations from your own domain; candidate, investor and press answers draw under 15%, pulling instead from review sites, forums, funding databases and news archives. To know what AI says about you, you have to test all four prompt sets separately.
That gap is the whole story. Everything your marketing team has done to shape AI product recommendations — documentation, comparison pages, structured data, review-site presence — barely touches the answer a recruit or a reporter receives.
The fastest way to check right now (10 minutes)
Before the methodology, the version you can run during a coffee break:
- Open ChatGPT, Gemini and Perplexity in logged-out or temporary sessions. Personalization otherwise feeds you your own history.
- Ask each engine four questions with your brand filled in: What does [brand] do? Is [brand] a good place to work? How much has [brand] raised? What are the main criticisms of [brand]?
- For each answer, note two things only: any factual claim that’s wrong or stale, and every domain cited.
Most teams find at least one wrong fact and discover that four to six domains they don’t control are doing the describing. That’s the gap this article is about.
How we measured the gap between buyer and non-buyer answers
We ran a tracking study inside MaxAEO across 40 B2B software and tech brands, from seed-stage startups to public companies, over a 30-day window in mid-2026.
The setup:
- 120 prompts per brand, split evenly into four sets: buyer/product-recommendation, candidate/employer-brand, investor/due-diligence, and journalist/media-background.
- Five engines: ChatGPT, Gemini, Perplexity, Claude, and Google AI Overviews.
- Prompts ran daily, in a clean session with no memory or personalization, from a US-based location.
- For every response we logged three things: whether the brand was named, the sentiment and framing of the mention, and every URL cited or linked in the answer.
That last field is what makes this comparable. Rather than arguing about which sources should matter, we counted which domains actually showed up in the citation list of 4,800 answers per brand-month.
Two rules kept the count honest: a citation counted once per answer no matter how many times it appeared, and answers where the engine declined to name any brand were logged as zero-mention rather than dropped.
Limits worth stating. This is a US-English, B2B tech sample. Consumer brands with heavy retail-review footprints and non-English markets behave differently — brands frequently rank well in one language and vanish in another, a split covered in maxaeo’s work on why a brand wins in English and disappears in German.
Where each audience’s answer actually comes from
Here is the citation mix, averaged across all 40 brands and five engines. "Overlap with buyer sources" is the share of domains in that prompt set that also appeared in the same brand’s buyer-prompt citations.
| Prompt set | Citations from your own domain | Top three external source types | Overlap with buyer-prompt sources |
|---|---|---|---|
| Buyer / product recommendation | 34% | Review platforms, comparison and "alternatives" blogs, vendor documentation | — |
| Candidate / employer brand | 9% | Glassdoor and Indeed, Reddit and Blind threads, job aggregators | 18% |
| Investor / due diligence | 12% | Funding databases, tech press, regulatory and filing sources | 22% |
| Journalist / media background | 7% | News archives, Wikipedia, press-release wires | 15% |
Read the last column first. Between 78% and 85% of the sources feeding non-buyer answers about your company never appear in a buyer answer at all. If your visibility program only tracks product prompts, you are watching one-fifth of the surface area that describes you.
Second thing to notice: your own domain contributes least where the stakes are most personal. On employer-brand questions, nine cited URLs in a hundred come from you. The careers page you rewrote last quarter competes with a two-year-old Reddit thread — and loses on volume.
This is a retrieval problem before it’s a content problem. Models chunk and rank passages by topical fit, so a question about interview loops surfaces passages that look like interview discussion, wherever they live. The practical consequence: you can’t win a source slot by publishing on a topic where you have no third-party footprint.
The four prompt sets, and what each one asks
Buyers: "Is this the right tool for me?"
The set most teams already track, so briefly: product prompts are the friendliest terrain you get. Own-domain citations peak at 34%, criticism appears in only 21% of answers, and comparison content does most of the work.
Two things teams still miss here. First, buyer prompts refine over several turns — a generic "best CRM" narrows to "best CRM for a 5-person team under $50 a month," and the brand that wins the broad prompt often loses the specific one. maxaeo’s breakdown of the refinement path from broad to specific prompts shows where that drop-off happens, and its research on how buyers actually phrase product-recommendation prompts covers the entry-level phrasings.
Second, the buyer relationship doesn’t end at purchase. Existing customers ask engines setup, integration and troubleshooting questions — and when documentation is thin, they get answers assembled from forum guesses, a support surface examined in how AI answers questions from customers you already won.
Candidates: "Is this a good place to work?"
Candidate research has moved from a Glassdoor tab to a chat window. In a May 2026 survey of 306 job seekers across seven countries, PerceptionX found 96% of AI-using candidates have researched an employer with these tools, 74% of them regularly, and 89% do it in ChatGPT. More than half — 58% — say they’ve hit inaccurate AI information about an employer.
The prompts that matter, from our tracked set:
- What’s it like to work at [brand]?
- Is [brand] a stable company to join in 2026?
- Has [brand] had layoffs recently?
- What’s the interview process at [brand] like?
- What does [brand] pay a senior engineer?
- [Brand] vs [competitor] — which is a better employer?
- Any red flags about working at [brand]?
Two patterns showed up consistently. First, compensation and layoff questions pull the thinnest sources — often a single forum post or an aggregator page with stale salary bands. Second, engines volunteer negatives on employer prompts that they suppress on product prompts: candidate-set answers surfaced a criticism or caveat in 47% of responses, against 21% for buyer-set answers.
The cause is framing. A buyer prompt reads as a recommendation request, and models hedge toward helpfulness. A "red flags" prompt explicitly requests criticism, and the model goes looking for it. maxaeo’s field notes on how AI answers employer-brand questions about your company unpack that asymmetry with example answers.
Investors: "What’s their traction, and what’s the risk?"
Investor prompts are the most factually specific — dates, dollar amounts, headcount, customer names — and therefore the most exposed to staleness.
- How much has [brand] raised, and from whom?
- What’s [brand]’s revenue or ARR?
- Who are [brand]’s biggest competitors and how are they positioned?
- Is [brand] profitable?
- What are the main risks with [brand]?
- Who founded [brand] and what’s their background?
- Has [brand] had any legal or regulatory issues?
In our sample, 31% of funding facts returned about the 40 brands were out of date by at least one round. Seven brands had raised a round no engine reported. Four had a Series A described as their latest raise when a Series B had closed more than six months prior. The lag concentrated in brands whose funding news ran only in their own newsroom and one trade outlet — thin coverage means slow propagation into the sources models rely on.
This matters beyond fundraising. Buyers on procurement committees run the same vendor-risk queries, and enterprise deals now routinely include an AI-assisted vetting step — the pattern covered in maxaeo’s walkthrough of how buyers use ChatGPT to vet your company before they talk to sales. The investor prompt set and the enterprise security-review prompt set are close to the same questions in different vocabulary.
Journalists: "Is this claim checkable, and who else has covered them?"
Reporters use these tools heavily. Muck Rack’s 2026 State of Journalism report, based on 897 responses collected between January and March 2026, found 82% of journalists use at least one AI tool, with ChatGPT at 47%.
Their prompts look different from everyone else’s — less "recommend," more "verify":
- What is [brand] known for?
- Who at [brand] speaks on [topic]?
- Has [brand] been in the news for anything negative?
- What are the main criticisms of [brand]?
- Which outlets have covered [brand]?
- Is [brand]’s claim about [X] accurate?
Journalist prompts had the lowest own-domain citation rate in our study at 7% and the highest reliance on news archives and Wikipedia. If your company lacks a Wikipedia entry and your coverage is confined to press-release wires, media-background answers about you tend to be short, hedged, and heavy on the single most-linked article about you — often your worst news cycle.
Bad news travels well in these answers for structural reasons: news content is dense, dated, entity-anchored and widely syndicated, which makes it both easy to retrieve and easy to rank.
One more reason the summary matters more than the sources behind it: across 27 markets, the Reuters Institute’s Digital News Report 2026 found just 4% of people always or often click through to the underlying sources in AI answers, versus 19% from search. Whatever the model compresses your story into is what lands.
Four failure modes that only show up in non-buyer prompts
Across the 40 brands, four problems appeared repeatedly — none visible in the buyer prompt set.
| Failure mode | Brands affected | What it looks like in the answer |
|---|---|---|
| Stale funding or headcount | 31% | Latest round or team size off by one or more updates |
| Single-source employer brand | 42% | One negative forum thread quoted as if representative |
| No named spokesperson | 55% | "Executives at [brand] have discussed…" with no attributable name |
| Competitor-framed identity | 27% | Company defined mainly by how a rival’s comparison page describes it |
The "no named spokesperson" number is the one most teams underestimate. In 55% of brands, no individual was reliably surfaced when the model was asked who speaks for the company on its core topic. That’s a direct loss for PR: a reporter asking "who at [brand] can talk about [topic]" gets a shrug, and calls someone else.
The competitor-framing failure deserves its own audit pass. When a rival’s "alternatives to [you]" page is one of the few crawlable pages describing your positioning in structured comparison language, it becomes disproportionately quotable — a dynamic detailed in maxaeo’s analysis of when a competitor’s alternatives page is AI’s main source about you.
What AI gets wrong most often, and why
Sorted by how often it appeared in our logs:
- Stale facts (most common). Funding rounds, headcount, pricing tiers, leadership. The model isn’t hallucinating — it’s quoting a source that was accurate 14 months ago. Fix the source, not the model.
- Over-weighted single sources. One Reddit thread or one review becomes "employees report…". Concentration is the tell: if one domain supplies most of your citations for a prompt set, that answer is one stale page away from breaking.
- Competitor framing. Your category, differentiators, and weaknesses described in a rival’s words because the rival wrote the only structured comparison.
- Dropped or hedged claims. Compliance, uptime and customer-count claims with no dated, linkable page behind them get softened to "reportedly" or omitted — the failure pattern maxaeo documents in making compliance claims that AI repeats correctly.
- Confusion with a same-named entity (least common, worst when it hits). Short or generic brand names get merged with an unrelated company. Distinct entity anchoring — consistent naming, an unambiguous About page, third-party profiles — is the only reliable fix.
How to audit all four prompt sets in an afternoon
- Build four prompt lists of 10–15 questions each — buyer, candidate, investor, journalist — using the examples above with your brand and category filled in. Write them the way a real person would type them, not in marketing language.
- Run each prompt in a fresh, logged-out session across at least ChatGPT, Gemini and Perplexity. Memory and personalization will otherwise feed you your own past queries.
- Log four fields per answer: mentioned or not, sentiment, every factual claim about you, and every cited URL. A spreadsheet works for a one-off; daily tracking needs an AI visibility tool that reruns the set and diffs the results.
- Mark every factual claim true, stale, or false. Stale is the biggest bucket in practice and the easiest to fix.
- Rank cited domains by frequency. The top ten domains across your non-buyer sets are your real reputation surface — treat that list as a work queue.
- Repeat the run 7 and 30 days later. Single-run results are noisy; engines vary answer to answer. Only repeated patterns are signal.
Budget roughly three hours for a first pass on 40 prompts across three engines. The output you want is one page: the ten domains that describe you to non-buyers, and the five claims that are wrong.
What to fix first, ranked by effort against impact
Start where a single correct source can overwrite a wrong answer. In our tracked set, factual corrections propagated into engine answers within 9 to 22 days when the corrected fact appeared in a source already cited for that brand — and often never propagated when published only on the brand’s own site.
In priority order:
- Fix the company-facts record where engines already look. Update your Crunchbase-class profiles and LinkedIn company page before rewriting your About page. Cited source beats owned source.
- Publish one attributable expert page per core topic. A named person, a bio, a topic, and a set of quotable positions. Cheapest fix for the 55% spokesperson gap, and it compounds across investor and journalist prompts.
- Answer the negative prompt honestly on your own domain. A page directly addressing pricing objections, limitations, or a past incident gives the model something authoritative to retrieve. It won’t outrank a forum thread on volume, but it changes the framing — the mechanism behind how AI handles late-funnel objection prompts about your brand.
- Feed the employer-brand sources rather than fighting them. Recent, specific reviews and an actively maintained profile on the platforms already being cited move a candidate answer faster than any careers-page rewrite.
- Make claims verifiable. Certifications, uptime, customer counts and compliance statements need a dated, linkable page behind them.
The ordering matters. Teams that start at step 3 — writing content — see slow movement, because they’re adding to the 7–12% of citations they already control. Teams that start at step 1 change the sources the model is actually reading.
How to measure whether it’s working
Track four numbers per prompt set, not one blended score:
- Mention rate — share of runs where you’re named at all. The floor metric.
- Factual accuracy rate — share of answers with zero stale or false claims about you. The only metric that matters for investor and journalist sets.
- Sentiment/framing split — recommended, neutrally listed, or caveated.
- Source concentration — how much of your answer depends on a single domain. Anything above 40% from one source is fragile; that domain going stale takes your answer with it.
Report these separately by audience. A brand can hold a strong buyer mention rate while its candidate answers are 60% negatively framed, and a blended "AI share of voice" number hides it completely. Watching accuracy drift over time also catches the slow failures — a funding fact that ages out, a review average that slides.
Frequently asked questions
How do I check what AI says about my company right now?
Open ChatGPT, Gemini and Perplexity in a logged-out or temporary session, and ask each one three questions: what your company does, whether it’s a good place to work, and what the main criticisms of it are. Log the answers and every source cited. That ten-minute version already shows you the gap between buyer and non-buyer framing.
Why do candidate and investor answers differ from product recommendation answers?
Because retrieval targets different content. Product prompts pull review platforms, comparison content and vendor documentation. Employer prompts pull review and forum sites; investor prompts pull funding databases and press. In our study, only 15–22% of the domains cited in non-buyer answers also appeared in buyer answers for the same brand.
Can I get AI to correct wrong information about my company?
You can change what it retrieves, which is what changes the answer. Correcting the fact on sources engines already cite for you — profile databases, review platforms, coverage — moved claims within 9 to 22 days in our tracking. Publishing the correction only on your own site frequently changed nothing, because that page wasn’t in the citation set to begin with.
Why does ChatGPT say something different from Gemini about my company?
Different retrieval indexes and different source weightings. In our runs, no two engines returned the same citation list for the same prompt, and disagreement was widest on employer and funding questions where source coverage is thin. Test at least three engines before concluding anything about "what AI says."
Do these non-buyer answers actually affect revenue?
Indirectly and unevenly, but the hiring-side effect is real: PerceptionX found 82% of AI-using candidates say AI research has shifted their perception of a company. On the buyer side, procurement and security reviewers run essentially the same risk and stability questions investors do, so a stale or hedged company-facts answer surfaces inside enterprise deals as well.
How often should I re-run these prompt sets?
Weekly is enough for stable categories; daily matters if you’re in a news cycle, mid-raise, or shipping a competitive positioning change. Single runs are noisy — engines return different phrasings for identical prompts — so treat any change as provisional until it holds for three consecutive runs.