Entity consistency across the web

This is the least interesting work in AI search and one of the most reliably useful, which is roughly the ratio for everything that functions.

It matters more now than it did five years ago because answers are assembled rather than listed. A list of ten blue links tolerates ambiguity, because the human sorts it out. A generated answer has to commit to a claim about a specific business, and ambiguity makes it commit to the wrong one, or skip you.

What is an entity, in practical terms?

The set of facts a system believes about one thing, and its confidence that those facts belong together.

For your business that includes the name, the address, the phone number, the category, the hours, the services, the people associated with it, the website, and every third-party page that mentions any of it. All of it is evidence, and it either agrees or it does not.

Disagreement is normal and it comes from ordinary business life. You changed your phone number. You moved suites in the same building. You shortened the trading name and never updated the old listings. A previous agency created a second profile. Your legal name and your sign have never matched.

Every one of those creates a fork, and forks split your evidence in half. Two half-documented businesses lose to one fully documented competitor.

What happens when a system cannot resolve you?

It guesses, confidently, and the failure mode is worse than being skipped.

Ahrefs ran the cleanest published experiment on this. It built a fictional luxury brand, verified the name returned zero Google results, generated 56 adversarial questions using a deliberately different model from the ones being tested, and put the brand in front of eight systems including GPT-4, GPT-5 Thinking, Claude, Gemini, Perplexity, Copilot, Grok and Google AI Mode.

In the first phase, with only the brand's own site published, the strongest models answered 53 or 54 of the 56 questions correctly. Perplexity failed roughly 40% of them, at one point insisting the brand made smartphones, having apparently resolved the unfamiliar name to a similar-sounding consumer electronics company.

That is entity resolution failing on a name alone, with no conflicting information anywhere on the internet. Now picture your business, which has three spellings, two phone numbers and an old suite number in circulation.

In the second phase Ahrefs seeded three mutually contradictory fake sources while leaving an official page on the brand's own site denying the claims. Perplexity and Grok repeated the fabrications as fact. Gemini and AI Mode flipped from skeptical to convinced. The brand's own denial lost to the third-party consensus. Ahrefs sells SEO software, so label it as vendor research, but the design is disclosed and it is a genuine controlled experiment.

The lesson for a real business is not about fakes. It is that you do not get to be the authority on your own facts by asserting them on your own site. You get there by having the rest of the web agree with you.

Where does consistency usually break?

Five places, in roughly this order.

Your name. Suffixes, ampersands, LLC, and descriptive tails added for keywords are the most common source of drift. Pick the name on your sign and use exactly that everywhere. Whitespark's 2026 survey of 47 local search experts still places keywords in the business title third among local pack factors at 223 points, which is precisely why people keep adding tails they should not.

Your address. Suite numbers, abbreviations, and the difference between the mailing address and the one customers walk into. Pick one written form and copy and paste it forever. Physical address in the city of search ranks fourth at 213 points.

Your phone. A tracking number on the website and a different number on the profile is the classic self-inflicted version. HTML name, address and phone matching the profile sits at fifteen on the same list at 153 points.

Your category. If you are described as a general contractor in one place and a remodeler in another, that is a genuine ambiguity about what you do, not a formatting problem. Primary category is the highest scoring individual factor on that survey at 227 points, and Darren Shaw's separate study of 1.8 million Business Profiles across 4,209 categories found that the exact match category outranks every adjacent one, with the pattern repeating for almost every category in the dataset.

Your description. Three sources say three different things about what you specialize in, usually because three different people wrote them years apart.

Grade the survey honestly. It is expert opinion rather than test data, and Shaw says so himself. The 1.8 million profile study is the harder evidence.

Older mentions cause most of the trouble, because nobody goes back to fix a 2019 roundup. You often cannot fix those, which is why the goal is a preponderance of consistent evidence rather than perfection.

What does one duplicate profile actually cost?

Here is the arithmetic. Use your own numbers.

The position. Your main profile carries 68 reviews, newest from last week. A duplicate profile a former agency created carries 19 reviews, newest from fourteen months ago. Your nearest competitor has 74 on one clean profile.

What the market sees. 68 against 74. You look behind. What you actually have is 87 against 74. You are ahead and losing on presentation.

Now the ongoing leak. You collect 5 new reviews a month. If one in five customers lands on the wrong listing, one review a month goes to the dead profile. Over a year that is 12 of your 60 new reviews, or 20%, spent on an asset you do not want to rank.

And the signal it sends. Recency of reviews is scored separately from review count, at eleven on the 2026 list. A profile whose newest review is fourteen months old reads as a business that may have closed.

Total cost of the duplicate: 19 reviews stranded, 20% of new review flow diverted, one profile actively signalling abandonment, and a name-and-address fork in your evidence base. That is why resolving duplicates is the highest-value item on the list and why it is worth the tedium.

Is this just NAP consistency with a new name?

It is broader, and the old version still holds inside it. Whether exact name, address and phone matching carries the weight it once did is a fair question, and the direction has not changed even if the tolerance has.

Shaw's summary of the 2026 shift is useful here: review and behavioural signals up, on-page and link signals down slightly, and citations losing about half a percentage point, continuing a downward trend. Directory citation building is fading as a lever. Entity coherence is not the same thing, and it is not fading.

The difference is scope. NAP consistency was three fields in directory listings. Entity consistency includes what you do, who you are, where you operate, what you are known for, and how third parties describe all of that in prose.

That prose layer is the part most businesses have never audited. The sources AI assistants cite for local queries are often small publications that described you once, from memory, three years ago. That description is now evidence.

It also matters more for local than for anything else, because local intent forces the system to resolve you to a physical place and a set of hours rather than to a topic. Topics tolerate fuzziness. Addresses do not.

This is another spot where the honest framing is that the work is not new, which is why so much of AEO is SEO with a new name. Fine by me. It just should not be priced as a new discipline.

Does schema markup fix this?

Partly, and the mechanism is worth getting right before you buy anything called an AI schema package.

Google states plainly that structured data is not required for its generative AI features and there is no special schema.org markup for them. That kills the direct pitch.

There is still a real reason to implement it. Fabrice Canel, a principal product manager at Microsoft Bing, confirmed at SMX Munich in March 2025 that schema markup helps Microsoft's language models understand web content. That supports the indirect path: search engines ingest structured data at index time to work out what your business is, and that resolved understanding feeds the answer layer. There is no credible published test showing a model reads your markup directly and cites you because of it. Two mechanisms, and vendors conflate them.

So implement it for entity resolution rather than for AI visibility. `LocalBusiness` plus the most specific subtype that fits, `Organization` with `sameAs` pointing at every verified profile you own, and `Service`. The `sameAs` list is the entity work, because it is the one place you formally state that these nine profiles are the same business.

One hard rule from Google travels with all of it: your structured data must match the visible text on the page. Marking up facts you do not display is a violation, not a shortcut.

How do you fix it without buying a subscription?

Manually, in an afternoon, and then never quite finish. That is the honest description.

Start with a source of truth. One document with your exact business name, address, phone, hours, primary category, secondary categories, and two sentences describing what you do. Everything else gets corrected against it.

Then search your business name in quotes and read everything that comes back, including the results that embarrass you. Old profiles, duplicate listings, dead pages, that directory you never signed up for. Log every version of your details you find.

Fix what you control first: your website, your Business Profile, your social profiles, the major review platforms. Then request corrections on the third-party pages that matter, which is a polite email and a surprisingly high success rate.

Duplicate profiles are the highest-value fix and the most tedious. Run the arithmetic above before you decide it can wait.

Two configuration traps worth checking while you are in there, both added as negative factors in the 2026 survey: a website URL that forwards to a different domain, and overlapping service areas across multiple profiles you own. Both are entity forks you created on purpose without meaning to.

You will notice entity problems in your own ranking data before you notice them anywhere else, because they tend to show as coverage that is patchy in ways that make no sense. Grid rank tracking rather than a single ranking number is how you see that pattern instead of averaging it away.

Your review responses count as public text too, written by you, describing your business. Consistency belongs there as well, though whether to respond to every review is a separate decision with its own tradeoffs.

What to do this week

1. Write the source-of-truth document. Name, address, phone, hours, categories, two-sentence description. One page. 2. Search your business name in quotes and log every mention with the details it asserts. 3. Fix your website and your Business Profile to match the document exactly, including punctuation. 4. Find and resolve duplicate profiles, and run the fork arithmetic so you know what the delay is costing. 5. Check your `sameAs` list against every profile you actually own. Most sites list three and the business has nine. 6. Email the three most-cited third-party pages with the correction and a thank you. People update more often than you expect.

Want to know where the web disagrees with you?

Send your business name, address and phone exactly as you would write them to eric@seod.com and I will check the places these details commonly drift and send you back the list of sources that disagree, plus which ones are worth fixing first. No charge, no pitch.

I offer this because the audit is tedious and mechanical and I am faster at it than you are, and duplicate profiles turn up often enough that it is worth the twenty minutes.

Otherwise, keep reading the AI search library.