On this page
What actually changed
For twenty-five years a search engine answered a question by handing you a list of places where the answer might be. The user did the last step: read the options, choose one, click. Everything an agency sold you was aimed at that click.
Several products now skip it. Google places an AI Overview above the results for a subset of queries, summarising an answer with a handful of linked sources beside it; AI Mode goes further and runs a whole conversational session inside Google. AI Overviews reached UK results in 2024 and AI Mode has followed. Alongside them sit ChatGPT with browsing and search, Perplexity, Gemini, and Copilot inside Windows and Edge. Each of them ends the journey with prose rather than a menu.
The consequence for a trade firm is specific and worth stating without drama. You can hold the same organic position you held last year and receive fewer visits from it, because the question that used to require your page has been answered above your page. Nobody has been demoted. The step where a human chose between options has simply been performed by software.
Two things have not changed, and both matter more than the noise suggests. First, queries with strong local commercial intent — somebody wanting a roofer in their town, today — still lean heavily on the map pack, the reviews and the phone number, because a written summary is a poor substitute for a business that answers. Second, these systems are still reading the open web to build their answers. They did not replace the index. They sit on top of one.
So the work has not been thrown away. A site that already ranks, loads quickly and says something concrete has most of what an answer engine needs. What it usually lacks is the last mile: being easy to extract from, easy to attribute, and corroborated somewhere other than its own homepage.
How a page gets chosen
Start with what is documented. Google has described AI Overviews and AI Mode as running a query fan-out: the system breaks your question into several related sub-questions, issues those against its search index, and generates an answer over what comes back. That is Google's own description of the mechanism, and it has a direct implication. Pages that cannot be found by conventional search cannot be retrieved by the thing built on top of conventional search.
The other engines work along broadly similar lines — retrieve first, write second — though the detail varies. Perplexity operates its own crawler and index. ChatGPT's search feature retrieves live results and cites them. Gemini grounds answers in Google Search. What none of them publish is the weighting: how a candidate passage is scored, how heavily prior reputation counts, what makes one of four equally accurate pages the one that gets named. That part is not publicly documented, and any agency describing it confidently is describing a guess.
What is observable, repeatedly, across engines, is a family of traits the cited pages tend to share. This is pattern-spotting rather than proof, and it should be read that way:
- The answer is present, not implied. A passage that states the thing directly can be lifted. A page that circles the subject for four paragraphs before arriving cannot be quoted without the model inventing the missing sentence, which it will generally decline to do in favour of a page that already said it.
- The claim has an owner. Figures attached to a publisher and a year survive retrieval better than floating numbers, because they can be attributed. This is also why our benchmarks page names the publisher, the year and the sample beside every figure rather than presenting a tidy table of averages.
- The same facts appear elsewhere. One site asserting something about itself is weak evidence. The same trading name, address, accreditations and service list appearing consistently across a Google Business Profile, Companies House, a trade body register and a supplier page is corroboration, and corroboration is what a system built to avoid confident errors is hunting for.
- The page is about one thing. Retrieval happens at the level of passages, not whole websites. A page covering six services in a general way gives a retriever nothing sharp to grip.
None of that is exotic. It is close to what a careful editor would ask for. The uncomfortable part is that most trade websites fail the first trait badly, because they were written to sound reassuring rather than to say anything a stranger could repeat.
The part that stops most firms dead
Here is the concrete technical point, and it is the one worth checking before you spend a penny on anything else on this page.
Most of the crawlers feeding these systems request your page, take the HTML they are handed, and stop. They do not behave like a browser. They do not generally download and execute your JavaScript and wait for the interface to assemble itself. Google's own crawler will eventually render, on a second pass that can lag; the answer-engine crawlers largely will not bother.
A great many small business sites built in the last few years are client-side applications. Request one of those pages the way a crawler does and what comes back is a near-empty shell: a container div, a bundle of scripts, and sometimes not even a heading. Every word of your copy, your prices, your service areas and your qualifications only exists after code has run. To a human with a phone the site looks immaculate. To something that reads HTML and leaves, it contains nothing at all.
You cannot optimise your way out of that. There is no content to rank, retrieve, quote or attribute. It is an invisible failure, because the site looks fine to everyone who checks it by looking at it.
Checking properly takes a minute. Open the page source rather than the inspector — the source is what was delivered, the inspector shows what JavaScript has since built — and search it for a sentence you know is on the page. If the sentence is not in there, that is the whole project until it is fixed.
This site is the worked example, and we will put it forward as one because it is the only claim on this page we can hand you the evidence for directly. pipelineos.co.uk is generated as static HTML and served that way. Every page carries its title, its description, its headings, its body copy and its structured data in the delivered source, before a single script runs. View the source of the page you are reading now and the paragraph you are reading is in it. That is not a clever trick and it is not new — it is how the web worked before it briefly stopped working that way — but it is exactly the condition an answer-engine crawler needs. We are applying the approach to ourselves, which is a different and smaller claim than saying it has been proven to produce citations. It has not been, by us or by anyone with data you can inspect.
Where a rebuild is genuinely warranted, that is a web design conversation, not an optimisation one, and we would rather say so than sell twelve months of content work into a site that cannot be read.
The work itself
Assuming the HTML is sound, the rest is unglamorous and largely cumulative.
Decide which crawlers you allow, deliberately
Your robots.txt is where you state this. The agents worth an explicit decision are GPTBot, OAI-SearchBot and ChatGPT-User from OpenAI, ClaudeBot, PerplexityBot, Google-Extended, Applebot-Extended, Amazonbot, Meta-ExternalAgent, Bytespider and CCBot. Two clarifications that get sold wrongly. Google-Extended and Applebot-Extended govern whether your content is used for those companies' generative training and grounding — they are not the switch that controls whether you appear in AI Overviews, which are built from the ordinary Google index via Googlebot. And robots.txt is a convention that well-behaved crawlers observe voluntarily; it is a request, not a lock. For most trade firms wanting to be found, the answer is to allow them all, but it should be a decision you made rather than a default you inherited.
Publish llms.txt, with realistic expectations
A proposed convention: a plain markdown file at the root listing your key pages with a line about each, so a model has a clean map of the site. It costs an hour. Be clear about its status — no major provider has publicly confirmed that it is consumed, and treating it as a ranking lever would be inventing a mechanism. We add it because it is cheap and harmless, not because it is known to work.
Mark up what the business is
Structured data describing your organisation, your services, your location and your questions and answers gives a machine an unambiguous statement of facts it would otherwise have to infer from prose. Google documents that structured data helps it understand a page; nobody has published evidence that it is a direct citation factor for generated answers. It is worth doing for reasons that were already good before any of this, which the glossary covers in plain English.
Answer in the first two sentences
Every page and every section opens by answering the question in its heading, then expands. Not a preamble about the importance of the topic. Number, range, condition, caveat. Then the detail for the human who wants it.
Shape headings like the questions people ask
“What does scaffolding cost for a two-storey house?” is a heading. “Our pricing philosophy” is not. This costs nothing and makes the retrieval job obvious.
Say the same things about yourself everywhere
One legal name, one trading name, one address format, one phone number, one list of accreditations, used identically on your site, your Google Business Profile, your directory entries and your trade body listing. Inconsistency here is the ordinary reason a system declines to state a fact about you: it found three versions and trusts none.
Attach dates and publishers to every figure
If you quote a number, name who produced it and when. If no reliable figure exists, say that instead of borrowing one. There is, for instance, no published independent UK cost-per-click benchmark broken down by industry, and saying so is more useful and more citable than repeating a vendor's average.
Keep it current, visibly
Stale pages lose to fresh ones on anything time-sensitive, and a review date on the page is a signal a human and a machine can both use. Our methodology page carries one for the same reason.
Read that list again and notice how much of it is simply good technical SEO with the vagueness removed. That is the honest summary of this discipline: it is not a separate channel with separate mechanics, it is the same foundations with a much lower tolerance for waffle. Which is why we treat it as an extension of SEO rather than a novel product with its own mythology.
What cannot be promised
This belongs early rather than buried, so here it is plainly.
There is no ranking dashboard for these systems. No equivalent of a rank tracker exists, because there is no stable ranking to track. Ask the same engine the same question twice and you can get two different answers with two different sources cited.
Results vary by user and by session. Personalisation, location, account history, which model version you happened to be served and plain generative variability all move the output. What you see is not what your customer in the next town sees.
Nobody can guarantee a citation. Not us, not anyone. The selection logic is not published, the output is not deterministic, and there is no submission process, no paid inclusion and no appeal. Anyone selling guaranteed ChatGPT rankings is selling something that cannot be verified even in principle — and the test is simple: ask them how you would check the claim independently at the end of the month. There is no answer to that question, which is why the promise exists.
We have taken the same line elsewhere on this site about lead-volume guarantees, and it is the same argument. A guarantee about a system whose output you cannot audit is not a guarantee. It is a sales device. The comparison pages go through how to interrogate that kind of claim before you sign anything.
The three honest measurements
You can measure something. You cannot measure the thing you would most like to measure. All three of these are partial, and knowing how they are partial is the point.
| Method | What it tells you | Where it falls short |
|---|---|---|
| Search Console | Google reports impressions and clicks from AI Overviews and AI Mode inside the Performance report, counted within web search totals. If your impressions hold while clicks fall on informational queries, that shape is informative. | It is folded into the totals rather than broken out, so you cannot isolate AI surfaces from ordinary results, and it says nothing at all about ChatGPT, Perplexity or Copilot. |
| Referral traffic | Visits arriving from chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com appear as referrals in analytics, and they are usually few but unusually well-informed, because the person arrived already briefed. | It counts only citations that produced a click. A citation that answered the question completely — often the most valuable kind — leaves no trace at all, and some traffic arrives with no referrer. |
| Manual prompt testing | Write down twenty questions a buyer would genuinely ask, run them across the engines on a fixed schedule, and record who gets named. Over months the pattern of who is being described as the obvious choice in your area is real information. | It is a sample of one, taken by one person, with no sampling frame and no way to know how representative it is. Treat it as observation, not measurement. |
Put together, those three give you a direction of travel and no precision. We will report them as exactly that. If what you need is a channel where attribution is clean and the arithmetic closes, that is paid search, where every click has a cost and a source, and where the UK market remains overwhelmingly concentrated — the IAB UK Digital Adspend 2025 study put total UK digital ad spend at £40.5bn with paid search accounting for £17.9bn of it. That is where the measurable money still is. This page is about something narrower and slower.
When this is the wrong spend
It is the wrong spend if your phone is not answered. Every route into a trade business, generated answer or otherwise, ends at somebody picking up, and the most common finding when we look at a firm's pipeline is a leak after the enquiry rather than a shortage before it.
It is the wrong spend if you need work this month. Nothing described here produces enquiries next week. Paid search does.
It is the wrong spend if your site cannot be read without JavaScript, because you would be paying for content nobody will retrieve. Fix the delivery first.
And it is the wrong spend if the honest version of the pitch does not satisfy you. The honest version is this: we can make your business easier to read, easier to quote and easier to corroborate, which is what the observable evidence suggests these systems favour, and we cannot tell you what share of answers will name you afterwards, because that number is not available to anyone. Our methodology sets out where our claims come from generally, the results page shows the four sectors where we have our own measured advertising data, and neither includes AI search, because we have nothing published to show you there yet. Saying so seemed better than implying otherwise.
If that is the arrangement you want, the two guides below go deeper on the practical side, the industry pages cover how enquiries behave in specific trades, the tools will do some of the arithmetic for you, and a twenty-minute call is usually enough to establish whether your site is even in a fit state to attempt this.
Want to know whether an AI engine can even read your site?The first check takes a minute: if your pages are built in the browser rather than sent as HTML, most AI crawlers see nothing at all. We will tell you which yours is.
Call 07443 392243WhatsAppCommon questions
Is AEO a different thing from SEO, or a rebrand?
Mostly the same foundations with less tolerance for vagueness. A page has to be crawlable, indexed and credible before any of it matters, which is ordinary SEO. What genuinely differs is the emphasis on extractability: direct answers near the top, question-shaped headings, figures with a named publisher, and consistent facts about your business across other people's websites. If someone is selling AEO as an entirely separate discipline with its own secret mechanics, be sceptical.
Will blocking AI crawlers protect my content?
It will stop the well-behaved ones fetching your pages, and it will also remove you from the answers they generate. Note too that Google-Extended does not control AI Overviews — those draw on the ordinary search index, so the only way out of them is to damage your normal search presence or use snippet controls. For a firm that wants to be recommended, blocking is usually the wrong instinct. For a publisher whose product is the text itself, it is a legitimate commercial decision.
How long before anything shows up?
Longer than paid search and on a less predictable schedule. Crawling and re-indexing take weeks, the models behind some of these products are updated on their own timetable, and corroboration across third-party sources accumulates over months. Anyone giving you a date for a first citation is inventing it, because nobody outside those companies can see the queue.
Do I need to write hundreds of pages to get cited?
No, and it tends to backfire. Mass-produced near-identical pages are what Google's scaled content abuse policy was written to catch, and a retriever has the same problem a reader does when every page says the same thing in a different order. A small number of pages that answer real questions specifically will outperform a large number that answer nothing.
Can you show me results from other clients doing this?
No, because we do not have any to show. Our published measured results come from scaffolding, roofing, glazing and surfacing advertising accounts, and we are not going to dress sector research up as a track record here or anywhere else. What we can show you is this website, which is built exactly the way we would build yours, and the reasoning on this page, which you are free to check against the documentation each company publishes.
Should this replace what I spend on Google Ads?
Not for most trade firms, and certainly not while the arithmetic on paid search still works for you. Treat this as a slow-compounding addition to the search side of the pipeline, funded out of the margin from channels that already pay, rather than a substitute for the one channel where you can still trace a pound to an enquiry.