DEV Community

Cover image for An AI assistant called this business "not one brand". Two weeks later it cited the founding year.
Canlah AI
Canlah AI

Posted on AI-assisted

An AI assistant called this business "not one brand". Two weeks later it cited the founding year.

Since we opened Canlah AI on 18 June 2026, many brands we meet ask the same thing first: are we in the AI's answer?

Here is one answer that changed, and what it taught us about why. We wrote this post with AI assistance; the figures come from our project records and our own report, linked at the end.

Before: found, but not recognised

An education business in Beijing. Someone asked Doubao, ByteDance's AI assistant, about it by name: is it any good?

Doubao called it "not one brand". It mixed the client up with several same-named schools in other cities and lines of business, and it hedged the whole answer with checks and warnings.

So the model had no trouble finding the name. What it couldn't do was tell which business the name belonged to.

After: two weeks later

Two weeks later, the answer laid out the founding story (running since 2017), three directly run campuses and four core strengths. The warnings were gone.

That is a client result taken from project records. It is one case, not a promised outcome.

Why a brand-name question still goes wrong

Asking by name is the easy case. In one 2026 preprint on ChatGPT, when the question named the brand, the model recalled it 99.4% of the time. In category questions that didn't name it, the same brand surfaced 3.32% of the time.

The Beijing case was that easy case, on a different assistant: the name was recalled, and the entity was wrong. A brand that shares its name with other schools, shops or apps can be "found" every time and still be described as someone else, or as nobody in particular.

Our report draws the line this way: work on your own pages mostly governs whether content that's already been retrieved gets cited, and what other sites publish about you governs whether it gets retrieved at all. Formatting alone (subheadings, Q&A blocks, schema) is hygiene, not leverage.

Same business, two weeks apart. The difference was what the AI could find.

Check yours in ten minutes

You don't need a tool to see your own starting point.

  1. Open a fresh chat with memory and personalisation off, so the answer isn't shaped by your history.
  2. Ask by name, twice: "What is ?" and "Is any good?" Add the city if you're local.
  3. Read the answer for three failure modes:
    • it merges you with a same-named business;
    • it hedges with checks and warnings instead of facts;
    • it misses the basics a buyer would want: when you started, where you are, what you do.
  4. Write down which sources it cites. Those pages are what the model is reading about you.
  5. Save the exact wording and re-run the same prompts on a schedule. Compare answers, not impressions.

If one of the three failure modes shows up, the fix starts with the pages the AI is reading, not with the AI.

Run it on a schedule

Canlah GEO, our Chrome extension, asks ChatGPT and Google AI Overviews the questions your buyers type, then shows whether the answer names you: canlah.ai/extension.


Sources: the Beijing case is a client result from Canlah AI project records (individual result, not a promised outcome). The 99.4% / 3.32% figures and the retrieval-versus-citation conclusions are from Marketing in the Agent Era (Canlah AI, 2026), canlah.ai/whitepaper/, which grades every outside figure by evidence tier.

Top comments (0)