What an answer engine reads before it names you

Between somebody asking and your name appearing there are a few seconds of reading. Knowing what gets read explains almost every odd answer you have seen.

The few seconds before the paragraph

Somebody types a sentence asking for a business like yours in your town. A few seconds later a paragraph appears naming two or three businesses.

Almost everything you can influence happens in those few seconds, and almost nothing about them is obvious from the outside. Once you know the sequence, most of the odd answers you have seen stop being mysterious.

It searches first

The most common misconception is that the assistant knows things about your business and is reciting them.

For local questions it mostly does not. It runs a search, reads what comes back, and writes from that. This is why answers about local businesses can include a listing corrected last week, and why they change between runs — the search step is not identical each time, and a great deal follows from that.

Some recall is involved. A widely discussed national brand may be described from what the model absorbed during training. A local business almost never is: there was not enough about you in the world for the model to have learned you as a fact, which is good news, because it means what gets read is current rather than frozen.

What comes back

A handful of pages, not the whole internet. Enough to write a paragraph from.

For a local question, that set is usually the large directories, a review platform or two, an aggregator specific to your trade, sometimes a community thread, and — if it has anything relevant to say — your own site. The set is small, which is why being in it matters so much more than being somewhere on the wider web.

The engine then reads those pages the way a hurried researcher would. It is looking for assertions it can repeat: names, addresses, hours, categories, prices, areas, the occasional evaluative sentence.

What it does when sources disagree

It reconciles, and reconciliation has a cost you can see in the output.

Where sources agree, you get a confident sentence. Where they disagree, you get either the majority view stated confidently — which is how a wrong hour survives a correction on your own site — or a hedge, telling the customer to ring and check.

The hedge is worth learning to recognise. An engine advising somebody to confirm your hours is frequently reporting a contradiction it found rather than an absence of information. That is a symptom with a specific cause and a specific fix.

The things it can lift without thinking

Some material is trivially usable and some requires interpretation, and the difference decides what gets quoted.

Trivially usable: a labelled field in a directory record, a heading followed by a plain sentence, a list, a table, a price with a currency symbol next to it, structured data. These are assertions with a subject and a value, ready to be dropped into a sentence with a source attached.

Requiring interpretation: a paragraph implying you have been established a long time, a photograph of a van in a recognisable place, a slogan, a testimonial that mentions a price in passing. A person infers all of this easily. An engine writing a sentence it has to stand behind mostly will not.

What it cannot read at all

This is where good businesses lose questions they should win.

Text inside an image — hours in a footer graphic, a price list as a picture, a service area drawn on a map — does not exist. A PDF price list is frequently not read. Content that only appears after a script runs may not be there when the page is fetched. A carousel rotating through your services may present one item to a crawler.

Every one of these looks complete to you, because you are looking at the rendered page in a browser. The check that matters is whether the fact exists as words in the document, and the fastest way to find out is to view the page source and search for it.

Why your best page can be invisible

A site can pass every technical check and still contribute nothing.

If your pages contain no checkable assertions — no price, no towns named, no specific description of a job — then there is nothing for an engine to lift, and it will build the answer entirely from directories. Your site was read. It simply had nothing to say that could be quoted.

That is the ordinary case, not an edge case, and it is the whole argument for making your own pages quotable.

Where the question is being asked from matters

Location is not a detail. For a local question it is close to the whole question.

The assistant knows roughly where the person is, and it searches accordingly. Two people in different parts of the same city can get different businesses named, and somebody a county away gets a different answer again — which is why an audit fixes the location deliberately, and why a friend in another town trying it for you proves very little.

It is also why claiming the towns you cover, in words, is worth more than it looks. You are not decorating a page. You are making yourself eligible for a question being asked from somewhere specific.

What a citation actually is

The word gets used loosely, so it is worth pinning down.

A citation is a source the engine drew on and is willing to show. Some interfaces list them plainly, some link them inline, some show them only when asked. They are not the same as links to you, and they are not an endorsement — a citation of a directory page about you is a citation, and so is a citation of a thread where somebody complained.

Two things make them the most useful artefact in this whole subject. They are observable, so you can see what your answer was built from rather than guessing. And they are addressable: unlike a ranking factor, a cited page is a specific document you can go and correct, claim, or improve.

That is why the practical advice keeps returning to them. Work the sources that appear, ignore the ones that never do, and treat the citation list as the shortest route to knowing what to fix. The mechanics are in the sources that keep coming back.

Reviews, and which parts get used

Reviews contribute differently to the rest, and understanding how changes what you ask customers for.

The rating supplies a number the engine can state. The volume supplies confidence — a handful of reviews is thin evidence, and engines tend to hedge rather than quote a rating that could swing on one more entry. The text supplies the adjectives: how you get described when the answer goes beyond naming you.

That last part is the one owners underuse. A review saying great service gives an engine nothing specific. A review naming the job, the town and what happened gives it a sentence it can use. You cannot script reviews and should not try, but asking somebody to mention what you actually did for them is a fair request, and it is worth more than another five-star with no words.

The order it reads in

Roughly: the question, then the location, then a search, then the sources, then a paragraph.

Each stage narrows the next. A question about a specific job in a specific town retrieves a different set of pages than a general one about the trade. The sources retrieved determine what can be said. And the paragraph is constrained by what those sources actually assert — which is why an engine that finds nothing specific about you writes a generic sentence, or writes about somebody else.

Reading the sequence backwards is the fastest way to diagnose a bad answer. If the paragraph is wrong, look at the sources. If the sources are wrong, look at the records. If the sources are fine and you are simply absent, the problem is upstream of all of it: nothing was retrieved about you, because nothing you published was worth retrieving.

What decides between two plausible businesses

When several businesses are credible candidates, the tie tends to break on three things.

Whether the sources agree about you, because a contradictory account is harder to summarise confidently than a consistent one. Whether there is anything specific to say, because a business the engine can describe precisely makes a better sentence than one it can only name. And whether somebody unaffiliated has said something about you, because that is the only evidence in the set that you did not write.

None of those is a trick, and none is available to buy. That is the honest and slightly dull core of this whole subject — what it is, in one page.

What it remembers versus what it looks up

One more distinction, because it explains a common piece of bad advice.

You will occasionally be told to get your business mentioned across as much of the web as possible so that the models learn about you. For a local business this is close to useless. The volume of material required before a model absorbs a fact about a specific small business is enormous, and the sort of low-quality mentions that can be manufactured at scale are exactly the ones that carry no weight.

What is looked up at question time, on the other hand, is a small set of pages that can be changed this week. That is where the leverage is, and it is why the advice in this library is repetitive about records and pages rather than exciting about reach.

Reading your own site the way it does

Twenty minutes, and it is the most clarifying exercise available.

Open your homepage source and search for the facts a customer asks about: a price, a town name, the word your customers use for the job. Then open your main service page and do the same. Anything you cannot find as text is not there.

Then ask an assistant your customers’ question and read what it cites. Between the two you will know exactly which half of the problem you have — whether the sources are wrong about you, or whether there was never anything to quote in the first place.