Most AI search optimisation advice is aimed at one stage of a six-stage process, and for most small businesses it is the wrong stage.
The advice says add statistics, add quotations, add FAQ markup. Those are citation tactics. They do nothing if the engine never ran a search, never crawled your page, or never pulled it into the handful of documents the model actually reads.
The clearest description of the full chain is in Olivier Martinez's critical survey of generative engine optimisation, published on arXiv in July 2026. It reviews 45 generative engine optimisation studies from late 2023 to mid 2026 and treats the path from prompt to citation as a pipeline in which every stage has to succeed before the next one matters. Limy's summary of the survey puts it in one line: passing one stage does not get you through the next.
The UK context makes this worth your time. Ofcom's 2025 Online Nation report says about 30% of searches now show an AI overview, 53% of adults say they see them often, and ChatGPT had 1.8 billion UK visits in the first eight months of 2025, up from 368 million a year earlier.
For each stage below: what controls it, what a UK SME can change, and one question that tells you whether this is where you are stuck.
The six stages and the levers under each
The diagram below is the survey's pipeline with the controllable levers attached. Stage 1 has no lever on your side at all.
flowchart TD
S1["1. Activation<br/>Does the engine search the web at all?"] --> S2["2. Crawling and indexing<br/>Can the engine fetch and store your page?"]
S2 --> S3["3. Retrieval<br/>Is your page in the candidate set?"]
S3 --> S4["4. Context allocation<br/>Does it make the shortlist the model reads?"]
S4 --> S5["5. Citation<br/>Is your page named in the answer?"]
S5 --> S6["6. Absorption<br/>Did your facts shape the answer?"]
L1["Lever: choice of query. Question-shaped prompts trigger AI answers far more often."] -.-> S1
L2["Levers: robots.txt per bot, text in HTML, indexed with a snippet, fast server"] -.-> S2
L3["Levers: topical relevance to each sub-question, classic SEO authority"] -.-> S3
L4["Levers: answer in the first lines of each section, sections that stand alone"] -.-> S4
L5["Levers: verifiable figures, named sources, liftable definitions"] -.-> S5
L6["Levers: be the origin of the fact, state it in one plain sentence, check fidelity"] -.-> S6
Stage 1: the engine decides whether to search
Nothing you publish matters until the engine decides to leave its own memory and look at the web. The survey calls this activation and places it with the engine, the query, the user's state and the time: "An engine may first decide whether to activate search."
The activation rates are lower than most people assume. Schulte et al, one of the reviewed studies, found that 57.8% of ChatGPT repetitions did not activate web search at all. For Google, Xu et al measured an AI Overview activation rate of 13.7% across their query set, rising to 64.7% for queries phrased as questions. Google's own documentation on AI features says overviews appear "only when our systems determine that it is additive to classic Search".
When the engine answers from memory, you are competing on what the model already believes about your market, which is slow to change. One 2026 preprint in the survey, by Sharma, found ChatGPT recognised 99.4% of products when they were named but surfaced them in only 3.32% of open discovery queries.
What you can influence here is the choice of battleground. Pick the prompts a real customer would type as a question ("which Northampton accountants handle R&D claims"), because those are the prompts that send the engine to the web. The two-word head term mostly gets answered from memory.
The question to ask: run your target prompt ten times in the engine you care about. How many times did it actually search? If the answer is rarely, no content change will fix this prompt.
Stage 2: crawling and indexing
If the engine does search, it can only use pages it can fetch and store, and this is where we see the most avoidable failures. Google's rule is simple: a page must be "indexed and eligible to be shown in Google Search with a snippet", and "there are no additional technical requirements".
OpenAI is different because it runs separate bots. Its crawler documentation lists GPTBot for training, OAI-SearchBot for search, and ChatGPT-User for fetches a person triggers inside a chat. "Each setting is independent of the others." Sites that disallow OAI-SearchBot "will not be shown in ChatGPT search answers".
If your site blocked everything with GPT in the name back in 2023, check what it still blocks. Blocking GPTBot is a legitimate choice about training. Blocking OAI-SearchBot removes you from ChatGPT's answers, which is a different decision that many owners never knowingly made.
Access failures are common even without a block. Allaham and Diakopoulos, in the survey, found 27.1% of the URLs that assistants pointed to could not be scraped because they were inaccessible, removed or non-textual. Google also warns that recrawling after a fix can take "several days to several months", so this is the stage to check first, not last.
The question to ask: open your robots.txt and your server logs. Does OAI-SearchBot appear in the last 30 days, and does Google's URL Inspection tool show the page as indexed with a snippet allowed? Our technical SEO work starts with exactly those two checks.
Stage 3: retrieval
Retrieval is where your page either enters the candidate set or does not, and no amount of citation polish rescues a page that was never retrieved. The survey: "If search is activated, the engine retrieves a set R and then ranks or reranks it." Google describes its version as query fan-out, issuing several related searches across subtopics and data sources.
Ranking in classic Google helps but is not the gate. Kirsten et al found that 53% of the domains Google AI Overviews linked to did not appear in the organic top 10, and that page overlap over two months was 18% for AI Overviews against 45% for organic results.
Fan-out is the opportunity for a small site. If the prompt is "how do I switch accountants mid-year", the engine may also search for notice periods and professional clearance letters. A page that answers one of those precisely can be retrieved without ranking for the head term.
There is a trap here that the citation-first advice never mentions. The SAGEO Arena study tested rewriting a page body for citation and found it reduced top-20 presence by about 9%, top-10 presence after reranking by 16%, and final citation by about 6%. Rewriting for the model made the page less relevant to the retriever.
The survey rates query-document relevance as the strongest reproducible factor in the whole corpus. In practice that means one page per real sub-question, in the words the customer uses, with the ordinary link authority SEO has always needed.
The question to ask: search the sub-questions, not the head term, in classic Google. Are you anywhere in the top 20 for any of them?
Stage 4: context allocation
Once retrieved, your page competes for a limited number of slots the model actually reads, and position inside that shortlist decides a lot. The survey is blunt about this:
"A perfectly formulated page that is absent from the top-k cannot contribute." Martinez, 2026
Vishwakarma et al ran 252,000 trials across six language models and eighteen factors, and found that relevance and position dominated which source was cited first. The original 2024 GEO paper by Aggarwal et al showed how far position swings things: under its "cite sources" method the fifth source gained 115.1% visibility while the first lost 30.3%.
That paper is also where the "GEO increases visibility by 40%" claim comes from. The survey traces it to a position-adjusted word count rising from 19.3 to 27.2 in a five-document test: a relative gain inside the context window, with no retrieval, clicks or sales measured. Treat anyone selling you a 40% uplift accordingly.
What you can influence is how your page reads when it is chunked. Engines rarely feed a whole page to the model. They take passages. Each section needs to answer its heading in the first sentence, name its own subject instead of leaning on "this" or "it" from the paragraph before, and carry its own evidence.
The question to ask: give the engine your page and the prompt together and ask it to answer from that page alone. Does it find the answer in the right section, or does it summarise your introduction?
Stage 5: citation
Citation is the stage nearly all AI search optimisation advice targets, and the evidence behind the popular tactics is thinner than the advice suggests. The survey notes that most GEO studies concentrate on the stretch between context allocation and citation.
The results are not encouraging for recipes. Puerto et al's C-SEO Bench found only three of 54 method-and-domain combinations produced a significant positive effect, and none in question answering. Keyword stuffing showed null or negative results. The survey's conclusion is that fixed recipes do not transfer well between engines and topics.
What did hold up, with moderate to strong support, was extractable evidence: verifiable figures, definitions and comparisons, provided they are truthful and attributed. Quotation addition was the strongest single method in the original GEO paper for the same reason. A model that is building an answer wants a sentence it can lift with a source attached.
So the lever is specificity you can defend. Your actual prices, your actual turnaround times, a definition written in one plain sentence, a comparison table with real numbers. Google's guidance is the same for its own features: no special markup, and structured data should match the visible text.
The question to ask: when the engine does cite someone for your prompt, look at what it lifted. A number? A definition? A step? Does your page contain that element in a form a model could quote?
Stage 6: absorption
Being cited and being used are different things, and the gap between them is where measurement usually goes wrong. The survey defines absorption as "the source's effective contribution to the facts, language, or structure of the answer", and fidelity as "the extent to which attributed claims are actually supported and rendered accurately".
Fidelity is worse than most people expect. Liu et al found only 51.5% of sentences in generative search answers were fully supported by their citations. Xu et al classified about 11% of 98,020 atomic claims in AI Overviews as insufficiently supported. You can be cited for something you never said, and your fact can appear with a competitor's link under it.
What you can influence is being the origin of the fact. Original data, your own pricing, your own documented process, instead of a restatement of someone else's. State each claim in one sentence that cannot be garbled, and keep it identical on every page that mentions it.
The question to ask: read ten answers that cite you. Is the claim attributed to you one you actually made, and is it correct?
Which stage you are actually optimising
Most SME budgets go on stage 5 tactics, where the evidence is weakest, while the gates at stages 2 and 3 go unchecked. A robots.txt line or a page that only ranks for the head term will cancel any amount of quotation-adding, and the SAGEO result shows citation rewrites can push you out at retrieval.
Two more findings from the survey shape how to measure. Answers are unstable: Schulte et al measured daily source-level overlap of 0.34 to 0.42 and suggest seven to eight repetitions per prompt before drawing a conclusion. And the link from citation to money is weak. The best traffic study found a 1.82 multiplier on referrals, but its placebo test failed, so the survey calls the effect suggestive and no more.
The survey's own summary line is the one to keep: "GEO is therefore a black-box optimization problem under incomplete information." You can see whether your page is crawlable, whether it is retrievable for the sub-questions, whether its sections answer first, and whether what gets attributed to you is true.
The decision rule we use is to work the stages in order and refuse to spend on a later stage until the earlier one passes. If a prompt never triggers a search, change the prompt. If the page is blocked or unindexed, fix that and wait. If the sub-questions never retrieve it, that is an SEO job. Only then does the citation work earn its budget.
If you would like the six checks run against your own prompts and pages, our AI visibility service does exactly that. We run them most for accountants and consultancies whose buyers now ask an engine before they ask around, or get in touch with the prompts you care about.