Does entity consistency actually change what AI answers say about you?

Because retrieval scores text against a query, and both families of scoring in use treat a rare exact string as special. Lexical scoring in the BM25 family weights a term by how rare it is across the collection, so a term almost nothing else uses carries most of the weight in a match.³ A product name is that kind of term. Split it into three surface forms and you have three separate rare terms, each present in a fraction of the passages, and a query using one form no longer reaches the passages that used another.

The dense half of a modern hybrid retriever is more forgiving of surface variation, and weakest exactly where you need it. The authors of Dense Passage Retrieval put the trade-off plainly: term-matching methods “are sensitive to highly selective keywords and phrases, but cannot capture lexical variations or semantic relationships well”, while the dense retriever “excels at semantic representation, but might lack sufficient capacity to represent salient phrases which appear rarely”.² A young product name is a salient phrase that appears rarely, the one case where neither half of the hybrid covers for the other.

Read that as a direction and not a magnitude. It says fragmentation should cost you something, not how much, and it assumes the engines run ordinary hybrid retrieval over web text, which their documentation supports in outline. The 2026 critical survey makes a neighbouring point from the measurement side, warning that parameters, redirects, AMP versions, translated pages and aggregators “can artificially fragment a domain”.⁴ Fragmentation is a documented problem in counting visibility. That it also reduces visibility is the step nobody has tested.

The load-bearing one is markup. The claim in its usual form is that schema markup improves your citation rate in AI answers, and Google rejects it in the guide it publishes for exactly this purpose: “Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add.” The same page tells you not to create “new machine readable files, AI text files, markup, or Markdown” for generative AI features, because Search does not use them.⁶ Both sentences were there at the page’s last update of 10 Jul 2026 and when this article checked it on 30 Aug 2026. Keep structured data for rich results if you already run it, and file it under ordinary SEO.

The second is the percentage. Entity optimisation is sold with effect sizes, and the effect size for this intervention does not exist in any study anyone can name. Where a number is quoted it usually traces to a vendor correlation between how often a brand is mentioned across the web and how often it turns up in AI answers, which measures mention volume rather than naming consistency, and correlates rather than tests. The coefficient belongs to a different variable.

The third is the inheritance. The phrase arrives from local search, where it attached to an explicit business record a vendor maintained and a merchant could edit. No AI answer engine publishes an equivalent record for a software product, so the move is an analogy rather than an inheritance, and the analogy is the part nobody argues.

Ask the same product two ways in a fresh session, on each engine you care about: first name it exactly and ask what it is, then describe the problem it solves in a buyer’s words without the name. The two failures look nothing alike. If the named run describes a different company, or attaches the wrong category with confidence, that is a resolution failure and naming work is the fix. If the named run is accurate and the unnamed run never mentions you, resolution is fine and you have a discovery problem, which naming work does not touch. That split is the one measured across 112 products, where the gap ran to roughly thirty to one on one engine.⁵

Then be honest about the sample. Five runs give a 95% Wilson interval of roughly ±33 points around a middling rate, which cannot tell 20% from 50%; thirty runs bring it to about ±17. Log what the survey’s minimum checklist asks of any study, because two runs differing on any of it are not comparable: “Product, mode, model, date, locale, account, and search enabled”.⁴

And accept what the test cannot do. It cannot tell you whether fixing the name changed anything, because there is no counterfactual to run and the variance between engines and between repeats is large enough to swallow an effect of this likely size. A before-and-after here is a story, not a measurement.