Somewhere this week, a buyer asked ChatGPT the exact question your company exists to answer. It gave them three names. Whether yours was one of them is a checkable fact, and most companies have never checked it.
The scale of the channel is not in dispute: OpenAI reported ChatGPT passing 800 million weekly users in late 2025, with growth continuing through 2026. Google says AI Overviews reach about 2 billion users a month. Perplexity reported handling roughly 780 million queries a month by mid-2025. Assistants still send fewer outbound clicks than Google search does, so the honest framing is not "AI replaced search" but "AI now shapes the shortlist." When a model names three vendors and you are not one of them, you lost the deal before your site got a visit.
How do you check what AI says about your company?
The method matters more than the tool, because the naive check lies to you. Rules first:
- Ask buyer questions, not your brand name. "Best fractional CTO services for SaaS" is the query that matters; "what is Acme Corp" only tells you the model can read your homepage.
- Use a clean session. A logged-in assistant that has seen your emails and your chats knows who you are and will parrot you back. Log out, use a fresh session, or query via API.
- Run each question more than once. Identical prompts produce different answers run to run. One appearance in three runs is very different from three in three.
Then pick the level of effort:
| Method | Cost | What you learn | Blind spot |
|---|---|---|---|
| Manual prompting (10 questions across ChatGPT, Gemini, Perplexity) | Free, ~1 hour | Whether you are named, who is named instead | Tedious to repeat; easy to bias the prompts you pick |
| Free share-of-voice check (e.g. Continuum's, HubSpot's AEO Grader) | Free, ~30 seconds | A structured score across engines, your gaps, competitors named instead | One-off snapshot, limited prompt count |
| Paid tracker (Otterly from $29/mo, Peec roughly $90-$500/mo) | $29-$500/mo | Daily monitoring, trends over time, citation sources | Measures the problem, does not fix it |
| Enterprise platform (Profound, Scrunch; custom pricing) | Five figures/yr | Deep dashboards, agent analytics, team workflows | Overkill below enterprise scale |
Disclosure: Continuum (this site) makes the free check in row two. It runs live buyer-intent questions for your space against ChatGPT, Gemini, Grok, and Brave and reports your share of voice with no signup: check your AI visibility. The paid trackers in rows three and four are compared in detail here.
Why doesn't AI recommend you even though you rank on Google?
Because ranking and being recommended are different gates, and the drop between them is steep. In an August 2026 first-hand study, we tested 45 queries a site ranked for on Google at position 12 or better. Live web search surfaced the site on 60 percent of queries. A browsing AI cited it on 15.6 percent. A non-browsing model, answering purely from training memory, named it 0 times in 9 attempts. Pages ranking as high as position 2.2 went uncited while lower-ranked pages were quoted.
So "we rank fine, we must be fine" is the expensive assumption. The model retrieves a pool of candidates, then quotes the clearest, most self-contained answer in it. If your page phrases its key facts as a pitch and a competitor's page phrases them as a neutral market figure, the competitor gets named, whatever the ranking order says.
How do you fix a bad answer?
Two jobs, matched to the two gates. Neither is fast, and anyone promising guaranteed AI placement is selling something nobody controls.
Make your pages liftable (the citation gate). Put a direct answer with real numbers in the first hundred words. Use question-shaped headings that match how buyers phrase the problem. Add comparison tables, because models lift finished structures. Phrase the market number neutrally before any product framing. This is ordinary editing work on your highest-intent pages, and it is the faster of the two jobs. The full mechanics are in the AEO vs SEO breakdown.
Earn mentions (the recall gate). Models without live search answer from training memory, which is built from years of mentions across the web. The 0-of-9 recall result above came from a site with good rankings and thin third-party mention density. Closing it means genuine roundup inclusions, real customer reviews, and being present where the buying question gets discussed. Months, not weeks, and authentic only; manufactured mentions risk the search rankings that feed the first gate.
How often should you re-check?
Monthly at minimum, and after every significant content change. Engines ship model and retrieval updates continuously, answers vary run to run, and a competitor reshaping their pages can displace you without your rankings moving at all. A free check makes the monthly cadence cost nothing; a paid tracker makes sense once the channel demonstrably drives pipeline.
The first step is not a subscription or a strategy. It is 30 seconds of finding out what the engines say right now: run the free check. If the answer is "invisible" and you want that fixed rather than watched, book a call and bring the questions your buyers actually ask.