The AI Visibility Guide

Part one · What actually changed

1. What changed

Search used to return a page of links and let you decide. Increasingly it returns an answer, or a short list of businesses with a recommendation attached. The moment the machine started making the shortlist, the job changed from ranking to being legible.

Google's own documentation is direct about this: there is no special markup that gets a page into an AI experience, and the guidance for AI features is the same guidance it has always given for search.^[1] That is worth sitting with, because most of what is being sold as AI optimization assumes the opposite.

2. Ranking and citation are not the same thing

Engines do not rank in the way a results page ranks. They synthesize an answer from sources and cite some of them. So the question is not what position you hold, it is how often you get cited, out of how many times somebody asked. Any number without that denominator is decoration.

3. Where the answer actually comes from

Here is what nobody tells you. When somebody asks an assistant for a contractor in their area, the answer gets assembled from whatever the system already holds about businesses like yours, plus whatever it retrieves at the moment of asking, and neither of those is your website alone. That gap between what you know about your business and what the machine knows is the Context Gap, and it is the thing every finding in this guide eventually traces back to.

• Your own site, if it can be crawled and if what it says is extractable
• Your Google Business Profile, which is a structured record rather than prose^[2]
• Directories and association listings that repeat your details, correctly or not
• Profiles you own but rarely think of as marketing, LinkedIn most of all
• Third-party mentions: forums, local news, supplier pages, community threads
• Reviews, and what they say rather than how many stars they average

4. Your website is a minority of the input

This is the finding most people are not ready for. In a typical measurement run, a business's own website is not the most cited source on the subject of that business. LinkedIn and community forums routinely place above it.

That is not a defect in our website. It is how the systems work. They weight sources they did not have to take your word for. Which means an AI visibility programme that only touches your website is addressing a minority of the input.

5. Engines disagree with each other

Engines disagree with each other far more than people expect. In a single run we have seen three engines cite a business at broadly similar rates while a fourth never cited it at all, despite returning hundreds of citations to other sources in the same run. It was not broken. It simply never picked them.

The practical lesson: never generalize from one engine. If somebody shows you a result from a single assistant, they have shown you a quarter of the picture at most.

6. The same prompt gives different answers

Ask the same question twice and you will often get different sources. This is why our method uses seven repetitions per query. One ask is an anecdote. Repetition is what turns it into a rate you can act on, and it is the single most common thing missing from AI visibility reporting.

Part two · What makes a business legible

7. The six pillars

Everything below sorts into six pillars, and they are not a framework invented so that we would have a framework; they are the six places a measurement keeps finding the problem.

8. Foundation: can a machine reach you at all

The least interesting work on this list, and the most frequently broken. A robots.txt that accidentally disallows the wrong path, a sitemap that has been returning a 404 since the last redesign, or canonical tags still pointing at a staging domain will each quietly remove you from consideration while your site looks perfectly healthy to every human who visits it.

1. Confirm robots.txt exists, returns 200, and does not block what you want read^[3]
2. Confirm sitemap.xml exists, returns 200, and lists the pages you actually care about^[4]
3. Confirm every page has one canonical, pointing at itself, on the domain you actually use^[5]
4. Confirm the AI crawlers you want are not blocked, and that this was a decision rather than a default

9. The Brief: saying what you do, for whom, and where

Context Before Content. A surprising number of service businesses never state their service area in text: it lives in the header image, or it is implied by the phone number, or it is mentioned once on a contact page that says almost nothing else. A machine cannot infer from a picture what you did not write down.

• Name the services in the words customers use, not the words your industry uses
• Name the area specifically. A city is better than a region, and a list of neighbourhoods is better again
• Name who it is for. A commercial general contractor and a residential remodeller are not interchangeable, and an engine cannot tell which you are from photographs
• Put it in text on the page, not only in an image, a video or a PDF

10. Liftable Facts: structured data, and what it does not do

Structured data describes your page in a format a machine does not have to interpret.^[6] For a service business the useful types are the organization, the local business and the services themselves.^[7][8]

What it does not do is win you anything on its own. Google is explicit that no additional markup is required to appear in AI features.^[1] Structured data removes ambiguity about facts that are already true and already on the page. It cannot manufacture a fact that is not there, and anyone selling schema as the unlock is selling you a label maker and calling it a factory.

11. The Paper Trail: proof a machine can verify

Systems weight sources they did not have to take your word for. That is the whole reason your own website places lower than you expect. The counter is to make your claims checkable somewhere other than your own marketing.

• Licence and registration numbers, written out, on a page that is crawlable
• Association and trade body memberships, listed where the association also lists you
• Named projects with locations and dates, rather than a gallery of untitled photographs
• Reviews that describe the work, which are worth more than reviews that only rate it
• Third-party coverage, including the small local kind nobody counts as press

12. Sounds Like You: voice that survives summarising

There is a real tension here, and it is worth naming before you rewrite anything. The structures that make a page easiest to extract are the same structures that make it read like every other page, and Google's guidance is to write for people first, saying plainly that content produced primarily to game a system is exactly what it is trying to filter out.^[9] Get this wrong and you start paying the Verification Tax: the hours that disappear checking, correcting and rewriting output that does not sound like you.

The resolution is not to choose. It is to keep your own headings, your own paragraphs and your own opinions, and let the machine-facing structure live in the markup rather than in the prose.

13. Everywhere Else: the surfaces you do not own

This is the opening. Most of the citation is decided out here, on surfaces you do not own and rarely look at, and almost nobody in the trades is working on them, which makes it the largest advantage available to a small business right now and the one that closes fastest once competitors notice.

1. Google Business Profile first, because it is structured and heavily weighted^[2]
2. LinkedIn second, because it is cited more often than most owners would believe
3. Directories and association listings third, and the goal is agreement rather than volume
4. Anywhere your business is described by someone else, corrected where it is wrong

If your profile says one thing and your website says another, the machine does not average them. It gets less confident about both.Jonathan Sterritt

Part three · How to measure it honestly

14. Why a citation rate needs a denominator

A citation rate is a fraction. Twelve citations means nothing at all until you know twelve out of how many asks, on which engines, across what period, and reporting the numerator on its own is the most common trick in this entire category. It is usually not even deliberate. It is just easier.

15. Writing the questions a buyer would actually ask

The measurement is only as good as the questions. The failure mode is writing questions that contain your own company name, which tests nothing more demanding than whether the engine can look you up when it is handed the answer, and that is the easy half. It is not the half that wins work.

• Ask the way a buyer asks, in their words, with no brand name in the question
• Include the location the way a person would say it, not the way a marketer would
• Cover the whole job, not only the service you most want to sell
• Include the awkward questions: cost, timeline, licensing, what goes wrong
• Keep the list stable, because a question you change is a question you can no longer compare

16. Repetition, and how many is enough

Because the same prompt gives different answers, a single ask tells you almost nothing. Our method uses seven repetitions per query, which is enough to separate a business that is genuinely in the consideration set from one that appeared once by chance.

If somebody cannot tell you how many repetitions they ran, they ran one. Fewer than five is noise dressed as a finding.

17. What a baseline actually contains

A baseline is not a score. It is a record complete enough that somebody else could repeat it and get a comparable answer.

• The exact questions, written out, unchanged
• The engines used, and the date range
• The number of repetitions per question
• The raw responses, kept, not summarised
• Citation counts with denominators, per engine, not averaged into one number
• What was already true about the site and the profiles on the day of measurement

18. The 90-day re-measure

Ninety days after the work lands, the identical measurement runs again. Same questions, same engines, same repetition count, same method. Anything else is not a comparison, it is a new measurement that happens to be more flattering.

This is also the part that protects you. A method that can show improvement can also show none, and knowing which is what you are actually paying for.

Part four · What is oversold

19. Most of the technical advice is overstated

There is no magic file that gets you into AI answers. Google's own documentation says no special markup is required for AI Overviews. Controlled tests of rewriting page bodies for machines have come back worse, not better. Treat anyone selling a secret technical unlock with the suspicion you would apply to any other shortcut.

What does hold up is unglamorous: a site that can actually be crawled, pages that clearly say what you do and where, facts a machine can lift and attribute, proof it can verify, and a consistent story across every profile you own.

20. The magic-file promise

Every few months a new file appears that is going to solve this. The pitch is always the same shape: add this one artefact and the assistants will understand you. Before you buy it, ask two questions. Which engines have publicly committed to reading it, and what happens to your visibility if they never do.

None of this makes such files harmful, and adding one costs about an hour. The harm is in the substitution, when a file replaces the boring work that actually decides the outcome.

21. The optimization tell

Worth knowing before you rewrite everything. The structures that maximise extraction, answer-first summary blocks and rows of parallel FAQ answers, are also the strongest signals that a page was machine-written. We had a piece scored 100% AI by a detector even though a human wrote every argument in it.

The way through is to keep the inline citations and the clean structure, which is where the value is, keep your own headings and paragraphs, which is where your voice is, and put the extraction scaffolding in the structured data where no detector is looking.

FAQ markup is a good example of the trade done well: the questions live in the markup where they can be read by a machine^[10], and the page itself keeps reading like something a person wrote.

22. Blocking AI crawlers

You can block them, and plenty of publishers have. Understand what you are buying: a business that blocks the crawlers is a business that cannot be recommended by the systems those crawlers feed.

For a publisher whose product is the words, that trade can make sense. For a contractor whose product is the work, it usually does not. Either way it should be a decision somebody made on purpose, and in our experience it almost never is.

Part five · What to do about it

23. What to actually do first

• Check you have a robots.txt and a sitemap.xml, and that they return 200 rather than 404
• Add structured data describing your organization, your location and your services
• Make sure your Google Business Profile agrees with your website on every detail
• Fix the profiles that are speaking for you, starting with LinkedIn
• Write down the twenty-five questions a buyer would actually ask, then measure what the engines say
• Measure before you change anything, or you will never know what worked

24. Fix contradictions before you add anything

Alignment Before Automation. The instinct is always to add: more pages, more content, more markup. The higher-yield move is almost always the opposite, because a machine reading three different service areas across three surfaces does not quietly pick the best one, it becomes less certain about all of them and hedges the answer accordingly.

1. List every surface that describes your business, including the ones you forgot you own
2. Write down what each one claims about services, area, hours and credentials
3. Find the disagreements. There will be more than you expect
4. Decide which version is true, then make every surface say that
5. Only then start adding

25. If you have no reviews yet

A common and fair objection: most of this assumes a paper trail you have not built. Start with what is already true rather than what you wish were true.

• Ask the last five clients you did good work for, individually, not by automation
• Ask them to describe the job rather than to rate you, because the description is the part that gets quoted
• Publish the jobs you have done with real locations and dates, even without a testimonial attached
• Get listed where your trade is listed, because a directory entry is verification you did not have to write

26. How to judge anyone selling you this

• Ask what their denominator is. If there is not one, walk
• Ask how many repetitions per query. Fewer than five is noise
• Ask which engines. Fewer than three is a sample, not a measurement
• Ask what they measured before they started. No baseline means no proof later
• Ask what they will not promise. Anyone promising rankings or citation rates is guessing

One more, and it is the one that separates most vendors: ask them to show you a case where the measurement came back flat. Anyone who has run this honestly has one.

27. A glossary, because half the confusion is vocabulary