An Assistant Is Telling Customers the Wrong Thing

A customer arrives annoyed. They asked an assistant whether you open on Sundays, it said yes, they drove over, you were shut. Or it quoted a price you stopped charging two years ago, or said you ship to a country you have never shipped to. You did not write any of that and you cannot see where it came from.

The instinct is to treat this as a hallucination and conclude nothing can be done. Sometimes that is right. Far more often the model read something true about a page that is out of date, and the wrong answer is a correct reading of a stale source.

Where does a wrong answer actually come from?

From one of five places, and they need different fixes. Working out which one you are dealing with takes ten minutes and saves a pointless complaint.

Your own live pages come first, because they are the most likely and the easiest to fix. Then your business listing, which you may not have touched in years. Then third party directories and marketplaces that copied your details once. Then old pages of your own that still resolve, which is the one that catches people. Last, and least often, the model's own training, which you cannot edit at all.

The ordering matters because effort should follow probability. Sellers routinely start at the end of that list, writing to the AI company, and never check whether an archived page on their own domain still advertises 2023 opening hours.

Do you need special AI markup to be read correctly?

No, and Google says so in plain words. Google's documentation on AI features and your website states that you do not need to create new machine readable files, AI text files or markup to appear in these features, and that there is no special schema.org structured data you need to add.

That sentence is worth keeping somewhere you can find it, because a growing number of services are sold on the opposite premise. It does not mean structured data is useless, since ordinary product and organisation markup has helped search surfaces for a decade. It means there is no separate AI schema you are missing, and anybody telling you there is has not read the documentation.

We reached the same conclusion when we looked at whether a dedicated file for language models does anything measurable, in the evidence on whether an llms.txt file changes anything. The pattern repeats: a plausible sounding artefact appears, nobody who operates the systems asks for it, and it sells anyway.

What can you actually control?

How much of your page may be used, through controls that already existed for snippets. Google's robots meta tag specification is specific about which surfaces each rule reaches, and the detail rewards a careful read.

The nosnippet rule applies to all forms of search results, and the documentation names web search, Google Images, Discover, AI Overviews and AI Mode. It also states the rule will prevent the content from being used as a direct input for AI Overviews and AI Mode. The max-snippet rule limits how many characters may be used, names the same surfaces plus Assistant, and likewise limits how much may be used as a direct input.

Now the part almost nobody notices. The data-nosnippet attribute, which marks specific elements inside a page rather than the whole page, is documented as designating textual parts not to be used as a snippet, and the documentation does not extend that statement to AI Overviews and AI Mode the way it does for the other two rules. If you are relying on element level exclusion to keep a paragraph out of an AI answer, you are relying on something the documentation does not promise.

That last row is where most of the confusion lives. Google-Extended is about whether your content trains or grounds some of Google's other systems. It is not the switch that governs whether you show up in an AI Overview, and we went through why the training question and the search question are separable in deciding which AI crawlers to allow on a shop.

How do you fix your business listing?

Directly, if you own it, and through a suggestion if you do not. Google's help page on editing your Business Profile describes signing in to the account linked to the profile, opening the profile, choosing to edit it, and saving.

Two things about that process are worth knowing before you need them. Pending changes suggested by other people are surfaced for you to accept or reject, which means an unclaimed or unwatched profile can accumulate edits you never approved. And where the problem is somebody else's listing carrying your information, or a listing that is plainly misleading, Google's page on reporting a business on Maps is the route, with evidence such as a photo or a link to the correct source attached.

Claim the profile even if you never intend to post on it. An unclaimed profile is a document about your business that other people can edit, and assistants read it as authoritative because it is structured and well maintained by the platform. This is the same reason local visibility work pays off beyond maps, which we covered in how AI changes local search for a shop with a physical location.

What about assistants that are not Google?

They are separate systems with their own agents, and the controls are not shared. OpenAI's overview of its crawlers documents three distinct roles: a crawler for collecting public content that may be used to improve and train its models, a search crawler that fetches public pages so ChatGPT can surface and cite them, and a user triggered agent that acts as a proxy for a person browsing.

The important consequence is that these are independent decisions. You can allow the search crawler so you remain citable while disallowing the training crawler, and you do not have to choose between all and nothing. A shop that blocked everything with an AI sounding name in its robots file has often removed itself from the citations it wanted while changing nothing about training it already happened to.

What does Google say does help?

Ordinary technical hygiene, which is a duller answer than the market wants. The same AI features documentation states there are no additional technical requirements beyond being indexed and eligible to be shown with a snippet, and that while specific optimisation is not required for AI Overviews and AI Mode, existing fundamentals continue to be worthwhile.

Its own list is worth reading as a checklist against a wrong answer problem. Make sure crawling is allowed in robots.txt and by any CDN or hosting infrastructure. Make content findable through internal links. Provide a good page experience. Make sure important content is available in textual form. Support text with high quality images and video where applicable. Make sure your structured data matches the visible text on the page. Check that your Merchant Center and Business Profile information is up to date.

Two of those items are doing most of the work for the problem in this article. Structured data that disagrees with the visible text is a machine readable contradiction on your own page, and a model given two versions of a fact may well pick the one you no longer mean. And the Business Profile line appears in Google's own list of fundamentals, which is a reasonable answer to anybody who thinks a listing is a marketing nicety.

There is also a caution in that documentation that applies to every promise made in this area. Meeting the requirements and best practices does not mean Google will crawl, index or serve the content, since indexing and serving are not guaranteed. Anybody selling certainty about appearing in an AI answer is selling something the platform explicitly declines to offer.

What if the source is somebody else's listing?

Then you are doing data entry across platforms, and consistency matters more than completeness. Aggregators, directories and marketplaces copied your details at some point and many have never rechecked them. A phone number you stopped using or a suite number that changed can survive for years in a dozen places.

The practical approach is to pick the canonical version of every fact once, write it down, and then make every platform match it exactly, including punctuation and abbreviation. A model comparing three sources that disagree has no way to know which is current, so it may take the most frequently repeated version rather than the most recent. Repetition beats recency when nothing carries a date, which is the opposite of how a person would judge it.

Start with the platforms customers actually use to reach you rather than trying to clear every directory. Your own site, your business listing, the two or three marketplaces you sell on, and whichever booking or delivery platform takes real orders. A long tail of scraped directories nobody visits contributes little and cannot be cleaned exhaustively.

Why do old pages cause so much of this?

Because a page that still returns a normal response is a live fact, regardless of how you think about it. The seasonal landing page from two winters ago, the price list you replaced but never removed, the shipping page for a courier you stopped using. Each is indexable and each says something about your business in the present tense.

A human visitor would notice the design looks old and the copyright says 2022. A model extracting a shipping threshold does not weigh those cues the way a person does. The fix is unglamorous: find the pages, then either update them, redirect them to the current equivalent, or remove them properly.

A quick way to find them is to search your own domain for the wrong fact rather than for the page. If an assistant claims you offer free delivery over 30, search your site for that figure. The page that states it is usually one you had forgotten existed, and the same audit tends to surface duplicate product pages that were quietly competing with each other.

How long does a correction take to show up?

Longer than feels reasonable, and the documentation is honest about it. Google's troubleshooting guidance for AI features tells site owners to confirm the control is actually visible to the crawler using the URL Inspection tool, then to allow time for recrawling, which it describes as taking from days to months, with the option to request recrawling.

Days to months is the sentence to quote to whoever is asking why it is not fixed yet. It also argues for fixing the source properly the first time rather than applying a partial patch, because each attempt costs another crawl cycle to verify.

Can you measure whether this is costing you anything?

Partly, and the honest answer is that the measurement is weaker than the problem. Google's documentation points site owners to Search Console to discover and diagnose technical issues, and performance for AI features is reported alongside the rest of search rather than as a separate channel you can isolate cleanly.

What that means in practice is that you will rarely be able to attribute a lost sale to a wrong answer. A customer who was told you close at four and did not come does not appear in any report. The absence is invisible, which is precisely why a quarterly manual check earns its place: it is the only instrument that sees the failure at all. We set out what the impressions reporting does and does not count in what the AI search impressions report actually tells you.

One number is worth watching as a proxy. Phone calls or messages asking to confirm something your site already states clearly tend to rise when a public answer contradicts you. If three people in a month ask whether you really ship to Ireland, something somewhere is saying you do not.

When is it genuinely not fixable?

When the claim exists nowhere you can edit, and it does happen. A model can blend two similar businesses, invent a plausible detail, or repeat something from a source that has since vanished. There is no edit button for a model's weights and no support queue that will retrain one for your opening hours.

What helps in that case is redundancy rather than argument. State the disputed fact clearly and consistently in several places a crawler reaches: the relevant page, your profile, your footer if it is something like delivery or hours. Models weigh consistent repetition across sources, so the practical defence against an invented fact is making the true one easy to find and hard to contradict.

It also helps to have somewhere authoritative to point a customer, which is an argument for a plain page of facts about your service that you control outright rather than a set of details scattered across platforms. That is one of the quieter benefits of running a storefront whose pages and code belong to you: when the facts need correcting, you are not waiting on anybody's release cycle.

A ten minute routine worth repeating

Once a quarter, ask two or three assistants the questions your customers actually ask. Your hours, your delivery terms, your returns window, whether you stock a specific thing. Write down the answers rather than reacting to them.

Then for every wrong answer, find the fact before you find fault. Search your own domain for the wrong figure. Open your business listing and read it as a stranger. Check your robots file for a block you added and forgot. In the large majority of cases one of those three turns up the culprit, and the thing you thought was a hallucination turns out to be your own site, telling the truth about a version of your business that no longer exists.