An AI visibility strategy is four decisions: which answers you intend to be named in, in what order you will earn them, what evidence you will accept that it is working, and when you will stop. Everything else is tactics. If your strategy document does not say what you are not doing this quarter, it is a wish list with a budget attached.
Most pages on this topic give you a list. Be clear. Be structured. Add examples. All true, all useless, because a list does not tell you what to do first when you have one person and £3,000 a month. This page is about sequence and constraint.
Why the checklist version fails
The advice you will find on this subject is usually six bullet points: clarity of offer, audience clarity, examples, steps and outcomes, internal linking, simple explanations. None of it is wrong. All of it is unactionable, for one reason. Every item is a quality standard, not a decision.
A strategy is defined by what it excludes. “Write clearly” excludes nothing. It cannot be scheduled, costed, or failed. Compare it with an actual strategic statement: we will earn citation share on the 14 prompts that precede a shortlist in commercial property finance, we will do it by fixing entity resolution then publishing against the six questions no competitor answers, we will accept nothing below a named appearance in ChatGPT and Google AI Overviews as evidence, and if we are not moving by month five we stop and put the money into paid. That can be argued with. That is the point.
Two things follow from taking it seriously.
First, you are choosing which questions to fight for. You cannot be named in every answer, and the businesses that try produce the thin near-duplicate pages that make a model less confident about them, not more. Scope is the first act of strategy.
Second, you are accepting a horizon. AI visibility work has a slow component you cannot buy your way past. Knowing that in month one prevents the month-four panic that kills most programmes just before they work.
Decision one: scope, which is a prompt list, not a keyword list
Keywords tell you what people type. Prompts tell you what they ask, which is longer, more conditional, and far more revealing about where they are in a decision.
Build the list like this.
Start with the questions your sales team is asked on first calls. Not the ones on your FAQ page. The ones that come up before anybody has agreed to a meeting. Those are the prompts your buyers are now typing into an assistant instead of asking you.
Add the comparison set. “X versus Y”, “alternatives to X”, “is X worth it”. Comparison prompts convert at a rate ordinary informational queries do not, because somebody running one has already decided to buy something and is deciding what.
Add the disqualifying questions. What would stop someone buying. Regulatory constraints, integration limits, minimum contract sizes. Answering these honestly is the single most reliable way to be cited, because almost nobody else does it and models reward the source that resolves the question rather than dodging it.
Then cut it to 20. Twenty prompts is enough to see movement and few enough to test monthly by hand. Our own prompt cluster research guide covers the mechanics of finding them.
Run those 20 across ChatGPT, Perplexity, Google AI Overviews, Gemini and Copilot. Record who gets named. That grid is your baseline, and without it nothing later in this page can be measured.
Decision two: sequence, and the order is not negotiable
There is a dependency chain here that most programmes ignore, and ignoring it is why money gets wasted. It runs identity, then extractability, then corroboration, then volume.
Phase 1, weeks 1 to 8: identity
A model has to be able to answer “who is this” with confidence before it will risk naming you in an answer. That confidence comes from agreement across independent sources. Your Companies House record, your website, your LinkedIn page, your directory listings and any press coverage should all say the same thing about your name, your address, your founding date and what you sell.
Most businesses fail this and do not know it. The failure is invisible from inside, because you know who you are.
Practical work: a single canonical description used everywhere, correct Organization schema, a public entity map so a model can resolve your presences to one organisation, and a robots.txt that actually admits the AI crawlers. That last one catches more businesses than you would expect. Blocking GPTBot and then paying for AI visibility is a self-inflicted wound.
This phase has a finish line, which almost nothing else here does. That makes it the right place to start even if you are unsure about the rest.
Phase 2, weeks 6 to 16: extractability
Now make the pages answerable. A model lifting an answer from your site needs a clean, complete, attributable passage. Long preamble, information split across three pages, and the actual answer living in a PDF are all extraction failures.
The test is blunt. Take your money page, delete everything except the first 60 words, and ask whether those 60 words answer the question the page is titled after. If not, the page is not citable, whatever its ranking. Our guide to content AI models actually cite goes through this properly.
Phase 3, months 3 to 9: corroboration
This is the slow one and it is the one that decides the outcome. A model weighs what other people say about you far more heavily than what you say about yourself. That means third-party mentions, independent reviews, sector publications, and the ordinary boring surfaces where a business’s existence is confirmed.
You cannot compress this with budget. You can start it earlier, which is the only lever available, and it is the reason it appears in this list at month three rather than month six. The difference between link building and citation building matters here, because the tactics that work for one are not the tactics that work for the other.
Phase 4, months 4 onwards: volume, and only now
Publishing at volume before phases 1 to 3 are done is the most common expensive mistake in this field. Forty articles attached to an entity a model cannot resolve produce traffic and no recommendations. The same 40 articles published after identity is fixed compound.
Decision three: evidence, agreed before you start
Decide in month one what would count as proof, and write it down, because in month four everybody will want to move the goalposts in whichever direction the numbers went.
| Signal | What it tells you | Honest limitation |
|---|---|---|
| Named appearance on tracked prompts | The actual product | Answers vary between runs. Test the same prompts, same phrasing, monthly |
| Share of the 20 prompts where you appear at all | Direction of travel | Slow. Expect nothing before month 3 |
| AI crawler hits in server logs | Whether the foundation layer works | Confirms access, not citation |
| Referral traffic from assistant hostnames in GA4 | Commercial reality | Under-reports badly. Many surfaces strip the referrer. A directional floor, not a count |
| Enquiries that mention an assistant | The only number a board cares about | Requires you to ask on the form. Add the field |
That last row is the one most programmes skip and then regret. Add “how did you hear about us” as a free-text field and read the answers. Our measurement guide covers the full metric set, and the GEO metrics page covers what to report upward.
Decision four: the stopping rule
Almost no strategy document contains one, which is why so much money gets spent on things nobody is willing to admit have failed.
Set it now. A reasonable one: if the 20-prompt grid has not moved by month 6, and crawler logs confirm the pages are being fetched, and entity resolution is verified clean, we stop and reallocate. All three conditions matter. Stopping because nothing moved when your robots.txt was blocking the crawlers the whole time is not learning anything.
For what the realistic curve looks like before you set that threshold, read when will you see AI citations.
What not to buy in the first 90 days
Say the uncomfortable thing, so here it is.
Enterprise AI tracking platforms. They are good products. They are also priced for portfolios and they will not change a single decision you make while you have 20 prompts you could test by hand in an afternoon. Buy one when the manual test becomes the bottleneck, not before. The comparison of tools against a managed programme is worth reading before you sign anything.
A second content supplier. More publishing into an unresolved entity multiplies the problem.
Anything sold with a citation guarantee. Nobody controls the output of a model they do not own. A guarantee here is either a misunderstanding or a sales tactic, and both are covered in red flags when hiring a GEO agency.
A rebrand. Every renaming resets the corroboration clock. If one is coming, do it before this programme starts, not during.
Where strategies differ by business type
The four decisions are universal. The weighting is not.
Regulated businesses front-load the evidence work, because a model applies a higher bar to YMYL topics and a compliance-safe published answer is a genuine competitive asset. See YMYL and AI search.
Local businesses get most of their movement from the off-site consistency layer rather than from publishing, because the recommendation is assembled from directory and review data. See AI search for local business.
Small businesses should compress phases 1 and 2 and skip phase 4 almost entirely, competing on specificity rather than volume. See AI search for small business.
Retailers have a product-data problem before they have a content problem. See AI search for retail.
Where you sit on the wider curve is worth checking against the GEO maturity model.
What a working strategy looks like on one page
If you take nothing else from this, take the shape.
- 20 prompts, chosen from real sales conversations, tested across 5 platforms, recorded.
- Entity fixed in 8 weeks. One description, one schema block, one entity map, crawlers let in.
- Top 10 pages made extractable, answer in the first 60 words, no exceptions.
- Corroboration started in month 3 and never stopped.
- Volume only after 1 to 3 are done.
- Monthly reading on the same grid, plus a source field on the enquiry form.
- A written stopping rule with three conditions, agreed before any money is spent.
That is the whole thing. It fits on a side of A4, which is roughly the length a strategy should be before it turns into a delivery plan.
MarGen runs this as the Synaptic Authority Engine, and the results are documented rather than described: a regulated B2B business from 0 to 37% citation share in 90 days and a law firm from 4% to 41% over 12 months. Pricing is published.
AI Visibility Strategy: Common Questions
What is an AI visibility strategy?
An AI visibility strategy is a set of decisions about which AI-generated answers you intend to be named in, in what order you will earn them, what evidence you will accept that it is working, and when you will stop. A list of tactics is not a strategy. If it does not tell you what you are not doing this quarter, it is a wish list.
How long does an AI visibility strategy take to show results?
Structural work can move within four to eight weeks, because fixing an entity a model cannot resolve removes something that was actively blocking you. Citation share moves on a three to six month horizon, because it depends on third parties publishing things you do not control. Anyone promising citations in 30 days is describing something outside their control.
Where should a strategy start if the budget only covers one thing?
Entity resolution. If a model cannot confidently work out who you are, every other investment is being spent on a brand the model is not sure exists. It is also the cheapest item on the list and the only one with a hard finish line, which makes it the one thing worth doing before you have decided anything else.
How do I know which AI platforms to target?
Run your 20 highest-value buying prompts across ChatGPT, Perplexity, Google AI Overviews, Gemini and Copilot, and record who gets named. Most businesses find they are absent everywhere or present on one surface only. Target the platform where your buyers actually are, which for UK B2B is usually ChatGPT and Google AI Overviews, and treat the rest as measurement rather than a workstream.
What should I not buy in the first three months?
Enterprise tracking platforms, a second content agency, and anything sold as an AI ranking guarantee. Tracking costs more than the decisions it will change while you still have fewer than 20 tracked prompts. A second content supplier multiplies an entity problem you have not fixed yet. Guarantees are a claim about a system nobody controls.
How much does an AI visibility strategy cost in the UK?
MarGen publishes its ladder rather than quoting case by case, and you can read the current figures on the packages page. Across the wider UK market only two of the seven agencies we track publish any pricing at all, which is worth knowing before you assume the quotes you are collecting are comparable.
Can I run an AI visibility strategy in-house?
The foundation layer, yes. Entity cleanup, schema, an llms.txt file and a monthly prompt test are all in reach of a competent marketer with a developer for a day. The part that is hard to do in-house is sustained third-party corroboration, because it needs relationships and a publishing cadence rather than knowledge.
What is the single most common strategy mistake?
Buying volume before fixing identity. Businesses commission 40 articles while their name, address, founding date and service description disagree across six public sources. The content gets crawled, the model cannot attach it to a confident entity, and the money produces traffic without recommendations.
Where to Go Next
Do the diagnosis first: free AI visibility audit · AI Visibility Score · how to audit your AI search visibility
Understand the underlying discipline: what is GEO · what is AI SEO · the 90-day AI SEO plan
Build the case internally: GEO ROI and the board case · in-house versus agency · 12 questions to ask a GEO agency
Have it delivered: MarGen is a UK GEO agency and AI SEO agency. The method is published, the pricing is published, and the results are in the case studies.