Entity and Brand Consistency

Search engines and language models both work with entities rather than strings. Before anything can rank or be cited for a brand, the system has to decide what that brand is, what category it belongs to, and which mentions across the web refer to it.

When the web describes a company inconsistently, that resolution fails or resolves wrong. The symptom is familiar: an assistant describes your product as something adjacent, attributes a competitor's feature to you, cites a three-year-old positioning line, or confuses you with a similarly named company in another industry.

None of that is fixed by writing more content. It's fixed by making every source agree.

Step 1: Audit what the web currently says

Do this before changing anything, and do it in a clean browser session with no personalisation.

1. Search the brand name. What's on page one? What does the knowledge panel say, if there is one? What autocompletes after the name?
2. Ask five assistants — ChatGPT, Claude, Perplexity, Gemini, Copilot — "what is [brand]" and "who makes [product]". Record the answers verbatim.
3. Search the brand name with the category and see whether you appear at all.
4. Search the brand name plus a competitor's name to see how you're positioned relative to them.
5. Check for collisions. Is there another company, a person, a product or a common word with the same name? That's the disambiguation problem, and it changes the strategy.

Write down every factual discrepancy. Those are the work items, and they're usually specific: an old tagline, a former product name, an outdated founder list, a category label that stopped being accurate two pivots ago.

Step 2: Write the canonical description

One paragraph, one sentence, and a fact block. Everything downstream copies from here.

NAME:        Exact legal and trading name, with capitalisation
CATEGORY:    The category, in words a stranger uses
ONE-LINER:   [Name] is a [category] that [does what] for [whom].
PARAGRAPH:   3-4 sentences. What it does, who for, what's distinct, founded when.
FOUNDED:     Year
FOUNDERS:    Full names
LOCATION:    City, country
WEBSITE:     Canonical URL, one form, with or without www — decide
LOGO:        Canonical file, one version

The one-liner is the load-bearing part. It's what gets quoted, and if every source says roughly the same sentence, that sentence becomes the answer. If every source says something different, the model averages them into mush or picks the loudest.

Be plain. "An AI-native SEO and GEO platform" resolves. "A next-generation growth intelligence layer" doesn't resolve to anything, because it isn't a category anyone recognises.

Step 3: Make the owned surfaces agree

Start with what you control, because it's free and it's the reference other sources copy from.

• Organization schema on every page, defined once with an @id and referenced. Name, URL, logo, description, founding date, founders, address, and sameAs. See schema-markup.
• sameAs is the key mechanism. It's an explicit statement that this entity is the same as the one at these other URLs. List every official profile: LinkedIn, GitHub, X, YouTube, Crunchbase, App Store, package registries, Wikidata if present.
• The about page. Real, substantial, with the founding story, the team, the location, the actual facts. Thin about pages are a genuine handicap for entity resolution.
• One name form everywhere. Not "Example", "Example Inc.", "example.io" and "Example App" depending on the page.
• Consistent title tag suffix across the site.
• A logo people can find, in a consistent form, at a stable URL.

Step 4: Make the off-site surfaces agree

This is the work, and it's tedious rather than difficult.

Update the description on all of them to match the canonical paragraph. Then keep a list, because a rebrand means doing this again and the list is what makes it a day rather than a month.

Step 5: Wikidata and Wikipedia

Wikidata is a structured, openly editable knowledge base that feeds knowledge graphs and is heavily represented in training data. Creating an item for a company requires the entity to be notable enough to have references, but the bar is much lower than Wikipedia's. If your company has press coverage, funding announcements or an established product, an item with accurate properties and real references is legitimate and useful.

Wikipedia is different. Do not create or edit an article about your own company. Conflict-of-interest editing is against policy, it gets reverted, and it can result in the article being deleted or the company being flagged. If the company is genuinely notable, someone else will write it. If it isn't, no amount of trying will make it stick.

Step 6: Corroboration

Consistency across sources you control gets you a coherent story. What makes a model confident in that story is independent sources saying the same thing.

• Press and industry coverage describing the company in the same category
• Genuine community discussion where the product is mentioned in context
• Documentation and integrations on partner sites
• Review platform presence with real reviews
• Conference talks, podcasts, open-source contributions under the brand

The thing not to do is manufacture this. Seeding forum mentions, buying listicle placements at scale and generating fake reviews are all recognised patterns, spam systems target them specifically, and a model trained on the wider web sees the inconsistency between the manufactured mentions and everything else.

Rebrands and name changes

The highest-risk moment for entity resolution, and the one most often handled as a design project.

1. Update every owned surface on the same day.
2. Redirect the old domain at URL level, permanently. See redirect-migration-plan.
3. Keep sameAs pointing at the new profiles, and keep the old profile pages live with a clear "formerly known as" note rather than deleting them.
4. Publish an announcement page explaining the change and keep it indefinitely. It's the document that lets a system connect the two names.
5. Update every third-party listing, starting with LinkedIn and Crunchbase.
6. Expect models to use the old name for a long time. Training data has a lag and there's nothing to do about it except be consistent everywhere from now on.

Measuring

Re-run the step 1 audit monthly and diff the answers. That's the whole measurement, and it's free for a single brand.

Doing it across five engines, multiple prompts and competitor comparisons every month is where it stops being practical by hand. Keupera runs it continuously through AI Brand Radar — visibility, position, sentiment and citations across ChatGPT, Perplexity, Google AI, Claude, Grok and Codex, benchmarked against competitors — over MCP:

claude mcp add --transport http keupera https://mcp.keupera.com/mcp -s user

Track two things over time: whether the description assistants give matches the canonical one-liner, and whether the brand appears at all for the category question. The second is the harder problem and the more valuable one.

Related skills

ai-visibility-audit, schema-markup, local-seo, eeat-author-pages, link-building-outreach, redirect-migration-plan