How do you explain an AI answer you disagree with to a client?

How do you respond when a client shows you a wrong AI answer about their business?

Ask to see the exact question they typed, and read the answer properly before saying anything about fixing it. Half the time the answer is defensible and the client is reacting to the framing. The other half it is genuinely wrong, and then you need a different conversation entirely.

The temptation is to react to their alarm with reassurance, which usually means promising a fix. That promise is the mistake, because you do not yet know whether the cause is your client's own pages, the wider web, or the engine assembling something nobody wrote.

These conversations have become a regular part of my work, and the honest version goes better than the confident one. Clients can handle "here is what we control and here is what we do not". They cannot handle a promise that quietly fails.

Why is promising a fix the wrong first move?

Because you are promising to change the output of a system you do not operate, based on inputs you only partly control. If the fix does not land, you have converted a problem the client had with an AI engine into a problem the client has with you.

I have watched people make this mistake in good faith. The client is upset, you want to be useful, so you say you will get it corrected. Three weeks later the answer has changed slightly or not at all, and now the trust conversation is about your competence rather than about a machine.

The better opening is diagnostic rather than reassuring. What exactly was asked, what exactly was returned, which sources were cited if any, and does our own site actually say something different. That takes twenty minutes and it is the only basis for any honest promise.

What can you actually control here?

Your own pages, and the degree to which they can be used. Google's documentation states that to limit the information shown from your pages in Search you use nosnippet, data-nosnippet, max-snippet, or noindex controls. That is the lever, and it is a restrictive lever rather than a corrective one.

Notice what those controls do and do not do. They limit how much of your content can be shown. They do not let you submit a correction, and they do not give you an editorial say over the sentence that gets generated. Restricting your own content is a blunt response to being described wrongly.

Google is also explicit that the access control sits at the crawl level. Its documentation says AI is built into Search and integral to how Search functions, which is why robots.txt directives for Googlebot is the control for site owners to manage access to how their sites are crawled for Search. There is no separate AI tap to turn.

Why do different engines say different things about the same company?

Because they are assembling rather than retrieving. Google's documentation describes AI features surfacing relevant links and using query fan-out techniques, which means one question becomes several underlying queries whose results get combined. Different fan-outs produce different combinations, and the same question asked twice can land differently.

So an inconsistency between two engines is not evidence that one of them is broken. It is the expected behaviour of systems that decompose a question differently and read a different slice of the web. Clients find this genuinely reassuring once it is explained, because it stops the wrong answer feeling like a verdict.

It also tells you where to look. If several engines independently say the same wrong thing, the cause is almost certainly out there on the web and probably on your client's own site. If only one says it, you are looking at that engine's particular assembly, and your leverage is much lower. Auditing this properly is the same exercise as auditing what AI engines say about a competitor.

How do you separate a wrong answer from an unflattering one?

Ask whether a careful journalist reading your client's own public pages could have written the same sentence. If yes, it is unflattering but defensible, and the work is editorial. If no, it is wrong, and the work is evidential rather than editorial, which is a different project entirely.

This distinction saves everyone's time, and clients rarely make it themselves. "They said we are best suited to small teams" often turns out to be a fair reading of a site that only shows small-team case studies. That is not an AI problem. That is the site working exactly as written, which is uncomfortable but fixable.

Genuinely wrong is different: a service you never offered, a location you are not in, a claim attributed to you that appears nowhere. Those deserve escalation. Unflattering deserves a rewrite. Treating the second as the first wastes months on the wrong remedy.

What should you tell a client about timelines?

That you do not know, and that anyone who gives them a number is guessing. I will not put a week count on when a corrected page changes what an engine says, because that depends on recrawling, on caching, and on how many other sources still say the old thing.

What I will commit to is a cadence of checking. Record the current answer with today's date, make the change, then re-ask the identical question on a schedule and log what comes back. That gives the client visible progress even when the answer has not moved yet.

The related trap is stale information the client has already fixed on their own site. That case is common enough that I have written about why AI answers about your company are out of date, and the honest expectation to set is that correction is a lagging process rather than an edit.

When is the wrong answer actually the client's own fault?

More often than anyone wants to hear, and saying so is the most valuable thing you can do. An old service page still live, an about page describing a previous business, a press release from a pivot ago. The engine did not invent anything; it read what is still published.

Google's documentation is worth quoting to a client at this point, because it removes the idea that some special technique is being withheld from them. It says there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary. The lever is the content itself.

So the conversation becomes a content audit rather than a technical escalation, which is both cheaper and more likely to work. It is the same discipline as fact-checking AI-written content before it goes on a client site: the source of truth has to be right before anything downstream can be.

What if the answer is genuinely damaging?

Document it immediately and completely, then escalate through the platform's own feedback route, and tell your client to take actual legal advice. I am not a lawyer and neither, probably, are you, and this is the point where being helpful means saying so.

Documentation means screenshots with visible dates, the exact prompt, which engine, and any cited sources, captured before anything changes. An answer you cannot reproduce is an answer nobody can act on, and these outputs are not stable, so evidence gathered later may simply not exist.

What I will not do is speculate to a client about what remedies they have or what a platform is obliged to do. That is outside my competence and the wrong person guessing about it can make a bad situation worse. My job is to fix what is on the site, capture the evidence properly, and be clear about where my role ends.

What should you do next?

The next time this lands in your inbox, resist answering for twenty minutes. Get the exact prompt, reproduce it yourself, capture it with a date, and check whether the client's own live pages could support the sentence. That sequence decides which of the conversations above you are actually having.

Then write the honest version of the update. Here is what the engine said, here is why it probably said it, here is what we control, here is what we are changing, and here is when we will look again. No promise about the output, a firm commitment about the work.

Across 350+ articles about how answer engines read pages, the thing I keep relearning is that the credibility you keep in these moments comes from what you refuse to promise. If a client has shown you an AI answer you cannot explain yet, reach out and let's work out which kind of problem it is before you reply to them.