Answers

How do you deal with AI systems getting facts about my brand wrong, or favouring older competitors?

You cannot edit a model's weights, and anyone offering to is selling something that does not exist. What you control is the evidence it retrieves when someone asks about you.

Two things fix most of it. The first is a canonical facts page — a plain-language statement of what your business is, what it does, what it does not do, and the terms it uses. Written for machines to read, not to rank. When a model has one unambiguous source for a fact, it stops inferring.

The second is consistency. Most invented details are not inventions at all: the model found two versions of your business and merged them. An old address on one directory, a name that shortened three years ago but not everywhere, a service you stopped offering. Reconcile those and the hallucination usually disappears, because it was never a hallucination — it was a contradiction.

On older competitors being favoured: that is usually corroboration, not age. They have more independent sources saying the same thing. The remedy is to build that evidence, which takes months and cannot be shortcut.

And then measure. I check what assistants actually say about a client rather than assuming, because the specific wrong belief is what tells you which source to fix.

Link to this answer

Which third-party sources will you actually target for us?

I do not answer this before measuring, and I would be cautious of anyone who does.

The sources that matter are different in every category, and they are knowable rather than guessable: when an assistant answers a question in your field, it cites its sources. That citation list is the target list. For one industry it is review platforms and a city directory; for another it is trade press, a professional association and a specific subreddit.

So the first piece of work is running your category's real buying questions and recording which domains produced the answers. Then the effort goes to those, in order, rather than to a generic list of fifty directories.

What is consistent across categories: sources you do not own carry more weight than your own website, because your site is you describing yourself. That is why corroboration is the slow, unglamorous heart of this work.

Link to this answer

What does the AI Authority Scan measure, and how will we prove it worked?

Four things, on a fixed set of prompts that does not change between runs.

Mention rate — of the buying questions in your category, how many produce an answer naming you. This is the headline number.

Per assistant, separately. ChatGPT, Gemini, Claude and Perplexity do not share an index and routinely disagree. A single blended "AI visibility score" hides the thing you need to know.

Sources cited. Which domains produced each answer. This is the diagnostic: it tells you where the work goes next.

Who appeared instead. The competitor set, which is often more informative than your own score.

Proof is the same prompts re-run on a schedule, compared against the first measurement. Not traffic, not rankings — the same questions, asked again.

One caution I give every client: these systems are not deterministic, and answers vary between runs. A small movement on a short prompt set is noise. Judge it on the trend across a fixed set, not on a single reading, and be suspicious of anyone reporting a precise score with no error range.

Link to this answer

How is optimising for AI crawlers different from optimising for Google?

Most of it is the same work, and anyone telling you otherwise is selling novelty. The foundations — clean structure, fast pages, accurate structured data, content that answers real questions — serve both. Roughly four fifths carries straight over from good SEO.

The differences that genuinely matter are these.

One entity, not many. Your business and you should be defined once, with every other page referencing that definition rather than restating it. Sites commonly declare the same organisation a dozen slightly different ways across their pages, which reads as a dozen similar organisations.

Crawler access is a decision, not a default. GPTBot, ClaudeBot, PerplexityBot and Google-Extended are separately controllable. Most sites inherit a robots.txt without ever deciding. If you want to be cited, allow them deliberately and document that you did.

JavaScript is the quiet killer. Google renders JavaScript; most AI crawlers do not, or do so shallowly. Content that appears only after a script runs is invisible to them. If your key facts arrive via JavaScript, you are invisible to the systems you are trying to reach — and your page will look perfectly fine to you.

Answer position inside the page. Models quote what is quotable. An answer in the first two sentences under a clear heading gets extracted; the same answer in the eighth paragraph does not.

Link to this answer

What happens to our visibility when the models update?

This is the right question to ask, and the answer tells you whether someone is building or gaming.

Everything I work on is retrieval infrastructure: consistent facts, independent corroboration, structured data, content that answers questions directly. Those are not exploits. They are what retrieval is built on, and a model update makes a system better at reading them, not worse.

Anything that works because of a quirk — keyword stuffing an llms.txt, manufacturing citations, prompt-injection tricks in page text — works until the quirk is fixed, and then costs you more than it earned. This field is young enough that those tactics are still being sold.

What does change is the measurement. When a model updates, the same prompts can produce different answers for reasons that have nothing to do with you. That is exactly why the benchmark is a fixed prompt set re-run on a schedule: it separates a model changing from your position changing. Without a baseline you cannot tell those apart, and you will either take credit or take blame for something you did not cause.

The honest summary: no one controls what an AI system recommends, including the people who build them. What survives updates is being genuinely, verifiably the right answer.

Link to this answer

Are you optimising for zero-click answers or for citation links that send traffic?

Both, but they are different outcomes and you should decide which one you are buying.

Most AI answers are zero-click and will stay that way. The assistant summarises, the user reads it, and nobody visits your site. Perplexity and the search modes of ChatGPT and Gemini do attach citation links, and those send real traffic, but they are the minority of interactions.

So if your only measure of success is referral traffic from AI, you will be disappointed by work that is actually succeeding. Being named in a zero-click answer is worth a great deal — it is the moment a stranger is told you are the right choice — but the visit that follows usually arrives as a branded search, a direct visit, or a phone call, and your analytics will attribute none of it to AI.

Practically that means two things. Optimise to be named, which is entity and corroboration work, and optimise to be clickable where links do appear, which is ordinary page quality. And measure branded search volume and direct enquiries alongside referral traffic, because that is where the return actually shows up.

Link to this answer

Do you optimise differently for informational versus transactional prompts?

Yes, and they need almost opposite work. This is the distinction most people in this field skip.

Informational — "what are the risks of X", "how does Y work" — is won with content. Clear explanations, answered in the opening sentences, structured so a passage can be lifted whole. Your own site can win these, because the model is looking for a good explanation and yours can be the best one.

Transactional — "best X near me", "who should I hire for Y" — is barely about your content at all. The model is not looking for an explanation, it is looking for a business it has confidence in. That comes from entity consistency, reviews, directory presence and third-party mentions. You can write the best page on the internet and still not be named.

Which matters more depends on what you sell. A practice that needs enquiries should put most of the budget on the transactional side, even though the informational side produces the nicer-looking content. Doing only content work and calling it AI visibility is the most common way money gets wasted here.

Link to this answer

How does this change if our customers are local?

Substantially, and it moves most of the work off your website.

Local answers are built from a different pool: Google Business Profile, mapping data, review platforms, local directories and geo-tagged mentions. Your website matters far less than it does for a national query. A local business with a mediocre site and excellent, consistent presence across those sources will be named ahead of a local business with a beautiful site and thin presence.

The practical consequences are that the single highest-return item is usually a complete, accurate Google Business Profile, that name, address and phone consistency stops being housekeeping and becomes the main event, and that reviews are a ranking input rather than a nice-to-have.

Measurement changes too. Local results vary by where the question is asked from — you can be named in one part of a city and absent two suburbs away — so a single test tells you very little. Anyone measuring local AI visibility with one query from one location is not measuring it.

Link to this answer

How do you stop us being lumped in with competitors in AI comparison tables?

You cannot control whether a model builds a comparison table or what it puts in one, and I would not trust anyone who says otherwise.

What you can control is what it has to say about you when it does. Two things determine that.

Whether your difference is stated as a fact. Models fill comparison cells from facts they can find. If your positioning lives in adjectives — premium, trusted, results-driven — there is nothing to put in the cell, and you get a generic row. If it lives in specifics — a named method, a defined deliverable, a documented outcome, a stated speciality — that is what appears.

Whether you appear at all. This matters more than placement. A business absent from the table has no position to defend. Getting into the comparison set is the first job; being described accurately within it is the second.

The uncomfortable part is that this is a positioning problem wearing a technical costume. If a model cannot tell you apart from three competitors, it is usually because the market cannot either. That is fixable, but not with schema.

Link to this answer

How much of the work happens on our website versus off it?

In most engagements the majority is off-site, and I would rather say that plainly before you engage than after.

On-site work — entity definition, structured data, content that answers real questions, crawler policy, making sure your facts are readable without JavaScript — is the faster, cheaper part. It is largely finite: it gets done and stays done, with maintenance.

Off-site work — directory presence, reviews, consistent facts everywhere you appear, mentions and citations from sources you do not own — is slower, harder, never finished, and carries more weight. A model trusts what other people say about you more than what you say about yourself, so this is where the durable gains are.

That has a scope consequence worth agreeing up front: who edits what. I work on your site directly or hand your developer exact changes, whichever you prefer. The off-site work needs access to your own profiles and, for reviews, your customers. Nobody can do that part without you, and a consultant who implies otherwise is either overpromising or planning to fabricate something.

Link to this answer