01Notes · 28 August 2026
Why Gemini Gets Your Company Wrong
Quick answer
Gemini blends retrieved sources with model knowledge. When it answers from memory rather than from a page, there is no citation to trace and no source to correct: the wrong fact lives in the model, not on a website.
That changes both what you can fix and how long it takes. Entity consistency matters more here than page-level ranking, and corrections land slower than anywhere else we measure.
Key takeaways
- Nobody can edit what a model says. Any agency claiming otherwise is describing something it does not control.
- What you can do is correct every source it might read, then re-sample until the corrected version wins.
- Corrections land slowest here. Weeks on Perplexity; considerably longer on Gemini.
- What gets recalled is an entity, not a URL. That is why entity consistency is the highest-yield work on this surface.
- Gemini is not AI Overviews. Different surface, sampled separately.
- Measure accuracy, not just presence. Being described wrongly costs more than being absent.
The difference between retrieved and recalled
Most AI visibility advice assumes a retrieval model: the system searches, finds pages, summarises them, cites them. Fix the pages and you fix the answer.
Gemini does that some of the time. The rest of the time it answers from what the model already contains — knowledge absorbed during training, with no page behind it and no citation attached.
That distinction decides everything about how you respond. A retrieved error is a source problem: find the page, correct it, done. A recalled error is not attached to any page you can edit. You cannot fix it directly at all.
Why your own site is probably inconsistent
Here is the part most companies have not checked. When a model recalls something about your business, it recalls an entity: a cluster of associated facts — rather than a specific page.
So the question is not "is my page correct". It is "do all the sources that describe me agree with each other". In our sampling they usually do not.
| Fact | Where it disagrees | Why it persists |
|---|---|---|
| Founding year | About page, LinkedIn, an old press release, Crunchbase | Nobody owns the historical record |
| Product tiers and pricing | Your pricing page versus comparison sites and directories | Third parties are never told when you change it |
| What you actually sell | Homepage versus a three-year-old category listing | Positioning moved; the listings did not |
| Markets served | Site copy versus old profiles naming markets you exited | Removal is nobody’s task |
| Leadership | Team page versus press coverage naming people who left | Coverage is permanent; employment is not |
Every row is a place where a model reading widely will find two answers and pick one. You do not get to choose which.
What you can actually do
Build a canonical fact set: the short list of things that must be true and identical everywhere, then correct every source you can reach. Your own site first, because it is free and fast, then third-party profiles, directories, comparison sites and anything else that describes you.
Then re-sample. Not once, and not next week. Corrections propagate slowly here, and a single reading of a non-deterministic system tells you nothing. That is entity optimisation, and Gemini is usually where the improvement shows last even when it shows most.
Set that expectation at the start of the work rather than in month three, because a client watching Gemini for movement in week six will conclude the programme failed when it is working normally.
Measuring a surface that often cites nothing
A meaningful share of Gemini answers carry no citation at all, which makes presence harder to evidence than on any other surface. Four things are worth recording:
- Cited or uncited. Whether the answer attributes anything.
- Named, cited, or absent. Coded separately, because named-without-link is common and invisible to analytics.
- Factual accuracy against your canonical fact set. This is the measure that moves most after entity work.
- Competitor set. Who appears alongside you, which tells you whether you are in the consideration set at all.
Any vendor giving you a single confident Gemini visibility number should be asked how they handled the uncited answers. How we handle it is published.
What this means for you
- Build the canonical fact set before anything else. Without it there is nothing to code accuracy against.
- Expect Gemini to confirm last. No movement in month two is normal, not failure.
- Refuse any claim to "fix what the model says". Sources can be corrected; model output cannot be edited by anyone.
02Common questions
Questions this raises
Why does Gemini state wrong facts about my company?
Because it blends retrieved sources with model knowledge. When it answers from memory rather than from a page, there is no citation to trace and no source to correct: the wrong fact is in the model, not on a website.
Can you make Gemini correct itself?
Nobody can edit what a model says. What you can do is correct every source it might read and re-sample until the corrected version wins. That works, and it is slower than on retrieval-led platforms.
How long do corrections take here?
Longer than anywhere else we measure. On Perplexity a corrected fact often appears within weeks. On Gemini it can take considerably longer, because the wrong version may have been absorbed during training rather than retrieved.
Why does entity work matter more on Gemini?
Because what gets recalled from memory is an entity, not a URL. If your founding year, product tiers or market coverage differ across your site, your press coverage and old directory profiles, the recalled version is unpredictable.
Is Gemini the same as AI Overviews?
No. AI Overviews are grounded in Google's index and ranking systems. Gemini blends that grounding with model knowledge and appears across several products with different behaviour.
What should I measure if there is no citation?
Four things: whether the answer cites anything at all, whether you are named, whether what it says is factually correct, and which competitors appear alongside you. Accuracy is the measure that moves most after entity work.
Sources
Every figure on this page comes from one of the following. Where a claim is ours rather than a source’s, it carries an evidence label instead.
- Rankability — State of AI Search, 1,645 citation observations across 31 topics, 2026. Source for the 24.1% cross-platform overlap and structural citation signals. Independent study
- Google Search Central — Optimizing your website for generative AI features (15 May 2026, updated 10 July 2026). Primary
- Searlixa sampling across client engagements, labelled Observed with sample and date. Our own
Ask an assistant
Hand this page to an assistant and ask it to summarise or challenge the argument. If it gets something wrong about us, that is itself a useful finding.
→Next step