AI Search Optimisation in 2026: The Complete Guide to GEO, AEO, E-E-A-T and Getting Cited by LLMs

Somewhere in the last eighteen months the job changed. Not the goal, because people still need to find you, but the mechanism. A growing share of the questions your customers ask now get answered by a machine that reads the web on their behalf, synthesises what it finds, and names three or four sources. You are one of those sources or you are absent.

This guide covers what actually changed, what the evidence supports, and what to do on Monday. It runs through generative engine optimisation, answer engine optimisation, E-E-A-T as Google currently defines it, and the mechanics of being cited by a large language model.

We have tried to be honest about the limits of what anyone knows. A great deal of what is published on this subject is stated with far more confidence than the underlying data supports.

First, the terminology mess

The industry has produced at least six names for overlapping ideas. Vendors have every reason to keep them distinct, because a fresh acronym sells a fresh product.

Term	What it usually means	Is it distinct?
GEO, generative engine optimisation	Getting cited inside AI generated answers	The broadest and most useful label
AEO, answer engine optimisation	Being the direct answer, including snippets and voice	Predates GEO and overlaps heavily
AIO, AI optimisation	Umbrella marketing term	Largely a synonym
LLMO, LLM optimisation	Optimising for the models themselves	Synonym
AISO or ASO, AI search optimisation	Vendor specific framing	Synonym
SEO	Traditional organic search	Genuinely distinct, and still the foundation

Our position is that there is one discipline here rather than five, and it has two halves. Making your content retrievable and extractable by machines. Making your organisation credible enough to be worth citing. Everything below sits under one of those two headings, and any acronym that fails to map to them is marketing.

How AI search actually works

Most advice fails because it treats an answer engine as a slightly different search engine. The retrieval pipeline is structurally different, and understanding it explains almost every tactic that follows.

Stage one, query fan out

The model does not search for the question it was asked. It decomposes it. Somebody asking which CRM suits a twelve person agency that bills hourly and needs Xero integration does not generate one search. It generates several. Best CRM small agency. CRM Xero integration. CRM hourly billing. Each one runs separately.

This is the most under appreciated fact in the field. You are competing for the short ordinary fragments the machine invents, rather than the long specific question the user typed. A page engineered around one elaborate query may never surface, because that query is never actually searched.

Stage two, retrieval

Candidate passages get pulled from a live index, the model’s training data, or both. Most systems use retrieval augmented generation, which means specific passages are fetched and handed to the model as working context rather than relying on what it memorised. The unit of retrieval is the passage, not the page, which is why document structure matters far more than it used to.

Stage three, synthesis

The model merges those passages into one narrative. It rewrites rather than quotes. There is an uncomfortable implication here. Your carefully worded positioning can be paraphrased into something you would never have written, and you have no editorial control over the result.

Stage four, citation

Sources get attributed and a fraction of users click. Citation counts are modest and differ by platform. Perplexity is the most generous at roughly five to ten sources per answer. ChatGPT search commonly cites three to five. Google’s AI Overviews usually surface three or four.

The arithmetic is brutal. Page one of Google offered ten organic slots. An AI answer offers three or four.

There is no position one

Language models are non deterministic. Ask the same question five times and you may get five different answers citing five different sets of sources. That breaks the mental model SEO has run on for twenty years.

There is no rank to hold. What exists instead is a frequency. Across many phrasings of many related questions, how often does your brand appear? The industry has settled on share of voice for this, and it should replace average position on your reporting.

Single observations are therefore worthless. When somebody tells you they checked ChatGPT and you were not there, that is one sample from a probabilistic system. You need dozens of prompts repeated over time before a signal separates from noise.

The numbers, and how carefully to hold them

Figures circulate freely here, usually stripped of methodology. These are the ones quoted most often in 2026. Most are vendor published and few have been independently replicated.

• Zero click searches rose from roughly 56 per cent in 2024 to about 69 per cent in 2025, so most searches now end without a click to any website.
• ChatGPT reports somewhere near 800 million weekly users.
• Queries put to AI assistants average around 23 words, against roughly 4 words for a typical Google search.
• Sessions run to about six minutes rather than seconds.
• Content older than about three months appears to see a marked drop in citation frequency.
• Pages carrying lists, statistics and quotable passages have been reported as 30 to 40 per cent more visible in AI answers.

The contradiction nobody wants to discuss

Here we depart from every other guide on this subject. The published data directly conflicts on how much AI citations overlap with Google’s top results.

• One widely cited analysis puts the overlap at roughly 60 to 70 per cent, which would mean ranking well on Google gets you most of the way to being cited.
• Research from the GEO firm Brandlight reports that the same overlap has fallen from around 70 per cent to below 20 per cent, which would mean Google rank is becoming a poor predictor of AI visibility.

Both were published in 2026 and they cannot both describe the same thing. The likeliest explanations are different query types, commercial against informational, different engines, or different time windows. The honest reading is that the answer is genuinely moving.

So hedge deliberately. Keep investing in classical SEO, because in the world where overlap stays high it is doing most of the work. Stop treating rank as a proxy for AI visibility, and start measuring citations directly, because in the world where overlap has collapsed rank will tell you nothing until revenue does.

Part one, make your content retrievable

None of the sophisticated work matters if the crawler cannot read the page. In our experience this is where most sites quietly lose, and it is the cheapest thing to fix.

Check you are not blocking the crawlers

AI crawlers are distinct user agents and each has to be permitted individually. The ones that matter today include GPTBot and ChatGPT-User from OpenAI, ClaudeBot from Anthropic, PerplexityBot, Google-Extended, and Bingbot, which feeds Copilot and parts of ChatGPT.

Two traps catch people repeatedly. The first is a robots.txt written years ago with a broad disallow that now blocks agents nobody anticipated. The second is nastier. Cloudflare changed its default posture to block AI crawlers, so a site can be excluded without a single line of its own configuration changing. If you sit behind a CDN, check the CDN as well as the file.

Verify this empirically rather than by reading configuration. Grep your server logs for those user agents. If they are absent entirely, something upstream is refusing them.

Render on the server

AI crawlers are substantially less capable than Googlebot at executing JavaScript. Content that materialises client side after hydration frequently does not exist as far as they are concerned. If your pricing, specifications or key explanatory copy arrive via JavaScript, they may be invisible to precisely the systems you are trying to reach.

The same applies to anything hidden behind interaction. Tabs, accordions, read more toggles and modal dialogs all require a click, and a crawler does not click. If the text is absent from the HTML your server returns, treat it as absent altogether. Quick test. Disable JavaScript in your browser and reload. What survives is roughly what the machine sees.

Structured data and llms.txt

Schema markup still earns its keep, particularly Article, FAQPage, HowTo, Organization, Person and Product. It hands machines unambiguous facts instead of requiring inference.

One serious warning. Only mark up content that is genuinely visible on the page. Emitting FAQPage schema for questions a visitor cannot see is structured data spam and a documented route to a manual action. Schema describes the page. It does not supplement it.

The proposed llms.txt convention puts a plain text map of your important content at the domain root. Be clear eyed about it. Adoption by the major engines is unconfirmed and no credible evidence yet shows it drives citations. It costs an hour, so treat it as a cheap option rather than a strategy.

Part two, make your content extractable

Retrieval operates on passages. Your job is writing pages full of clean self contained passages a machine can lift without damaging them.

Answer first, then explain

Journalistic instinct says build toward a conclusion. Machine retrieval punishes that. Put the answer in the first forty to sixty words of the relevant section and elaborate afterwards. A definition that arrives in paragraph six is competing against every other page’s paragraph one.

Write passages that stand alone

This is the subtlest point in the guide and the one most often missed. A passage opening with “as we saw above, this approach” becomes useless the moment it is extracted, because “this approach” has lost its referent. Each section should survive on its own. Name the subject instead of relying on pronouns. Repeat the key term instead of writing “it”. Prose that reads slightly over explicit to a human reads as unambiguous to a machine.

Structure the document honestly

One <h1>. A logical <h2> and <h3> hierarchy with no skipped levels. One idea per section. Question shaped headings where a question is genuinely what somebody would ask. Paragraphs of two or three sentences. Tables for comparisons and lists for sequences.

None of that is decoration. Headings are the primary signal a retrieval system uses to work out what a chunk of text is about.

Target the fan out rather than the question

Take your most valuable long question and write down the three to five short queries a machine would plausibly derive from it. Then check honestly whether you have content that would be retrieved for each fragment. Usually you do not, and that gap is your content plan.

Part three, E-E-A-T from the source

Most GEO guides ignore E-E-A-T entirely, which is a mistake, because the credibility signals that make a page worth ranking are largely the ones that make it worth citing. So much writing on this topic paraphrases other paraphrases that the following comes straight from Google’s own published documentation.

Trust is the one that counts

Google states explicitly that of the four components, trust matters most. Experience, expertise and authoritativeness are contributors to trust rather than four equal boxes to tick, and content does not necessarily need all of them. A page can display abundant credentials and still be untrustworthy, and Google says that fails regardless.

The weighting rises sharply for Your Money or Your Life subjects, meaning anything that materially affects health, financial stability, safety or public welfare.

Who, how and why

Google frames self assessment around three questions, and they demand more than the checklists usually derived from them.

Who wrote it. Is authorship self evident, with bylines where a reader would expect them, linking to genuine background on the author’s expertise?

How it was made. For anything evaluative, show your working. How many things you tested, how you tested them, what the results were, with evidence such as photographs. This is the requirement most content quietly fails.

Why it exists. Where the primary purpose is attracting search traffic rather than helping somebody, Google states plainly that this is not what its systems aim to reward.

What Google actually says about AI written content

This gets misreported in both directions, so it is worth stating precisely. Using AI is not itself a violation. Using automation to generate content primarily to manipulate search rankings is a spam policy violation.

Google also asks publishers to consider disclosure where a reader would reasonably expect it, covering that automation was involved, how it was used and why it was useful. Note the framing. Disclosure is a question of reader expectation rather than a blanket rule.

The self assessment question that should worry most publishers is this one. If your content draws on other sources, does it avoid simply copying or rewriting them? A great deal of AI assisted content fails exactly there. So does a great deal of ultimate guide content assembled by hand from the top ten results.

Part four, authority beyond your own website

Your own site is one input among many. Models form a view of your organisation from the whole web, and that view often gets decided somewhere you do not control.

Unlinked mentions carry weight. Unlike classical link equity, a plain textual mention of your brand appears to contribute, which makes digital PR and community presence directly relevant.

Getting into the sources already being cited is the fastest route in. Find which pages the engines already quote for your target questions, then work to be represented there. Being added to an existing frequently cited comparison article beats building a competing page from nothing.

Community platforms punch above their weight. Reddit, YouTube and specialist forums appear disproportionately in AI answers. Genuine participation helps. Astroturfing is detectable and damaging.

Wikipedia matters more than seems reasonable. Some analyses attribute close to half of ChatGPT’s top citations to it. If your organisation genuinely meets notability requirements an accurate entry is valuable, though attempting to manufacture one fails badly and publicly.

Consistency is itself a signal. Name, address, founding date, leadership and descriptions should match everywhere. Contradictory facts across the web make a machine less confident about asserting anything at all about you.

Part five, freshness and the citation cliff

AI systems show a pronounced recency bias, considerably stronger than classical search. Multiple practitioners report citation frequency falling away sharply once content passes roughly three months old.

Two warnings. That observation is largely vendor reported and we have seen no independent replication, so treat it as a working hypothesis. More importantly, Google explicitly lists changing dates to appear fresh without substantive updates as a red flag. Rewriting the date is the thing Google names as a warning sign rather than a freshness strategy.

The defensible version is a genuine quarterly review of your most valuable pages. Update statistics, replace stale examples, add what has changed, remove what is now wrong. Worth doing on its own merits whatever the recency effect turns out to be.

Part six, entities and why the machine needs to know what you are

Classical SEO thinks in keywords. AI retrieval thinks in entities, meaning discrete things in the world with properties and relationships. Your company is an entity. So is each of your people, products, locations and areas of expertise. Before a model can recommend you it has to have formed a stable idea of what you are.

That explains a pattern which otherwise looks arbitrary. Two agencies with comparable content and comparable links, where one gets named constantly by AI engines and the other never does. Usually the difference is that one has a coherent corroborated entity across the web and the other has a website.

What builds an entity

Corroboration across independent sources does most of the work. One page saying you specialise in something is a claim. Fifteen unrelated sources saying it is a fact as far as a machine is concerned.

Structural consistency does the rest. Legal name, trading name, address, founding date, leadership. Where those disagree across directories, LinkedIn, Companies House and your own footer, confidence drops. Organization schema with a sameAs array pointing at verified profiles tells a machine these accounts are all the same entity rather than leaving it to infer.

Named people help disproportionately. Individuals are easier to establish as entities than companies and their credibility flows back to the organisation. A recognised specialist writing under their own name with a consistent presence beats an anonymous corporate byline.

The specificity trap

Most businesses describe themselves too broadly to be useful. “Full service digital marketing agency” describes a category rather than an entity, and gives a model no reason to select you over five thousand identical claims.

Entities get retrieved on specificity. “Paid search and marketing automation for UK aesthetics clinics” is a narrower claim and therefore a far stronger one, because when somebody asks about exactly that there is a defensible reason to name you. Breadth feels commercially safer. In retrieval it approaches invisibility.

Part seven, what kinds of content actually get cited

Some formats are structurally more citable, because they hand a synthesising model something it cannot manufacture for itself.

Original data is the strongest position available

A model generating an answer needs facts. Where a number exists in only one place and the answer requires that number, the source gets named. This is the one advantage nobody can compete away by writing a longer article, because they cannot obtain your data by writing.

It need not involve a research budget. A survey of two hundred customers, an analysis of your own campaign results across a year, benchmark figures from your client base, or a documented test with a stated method all qualify. What matters is that the number originates with you and the methodology is stated plainly enough to trust.

Definitions

A crisp self contained definition in the first two sentences under a heading is among the most extractable structures there is. “X is …” followed by one clarifying sentence can be retrieved verbatim. Own the definition of a term in your field and you get cited every time somebody asks what it means.

Comparisons and tables

Comparative questions make up a large share of commercial AI queries, and a well built table answers them in a form requiring almost no synthesis. Tables also survive extraction better than prose, because the relationships sit in the structure rather than being implied by sentence order.

One caveat. Comparisons that conveniently conclude you are best get recognised by readers and models alike. A comparison conceding where a competitor is genuinely stronger is more likely to be trusted and more likely to be cited.

First hand accounts

This is the experience in E-E-A-T and the hardest thing to fake. What actually happened when you ran the thing. What broke. What the numbers were before and after. What you would do differently. No amount of summarising other people’s articles produces this, which is exactly why it is valuable.

Part eight, internal linking and topical depth

Because engines fan queries out into fragments, coverage across a topic matters more than the quality of any single page. A site with one excellent article and nothing around it gets retrieved for one fragment. A site with a coherent cluster gets retrieved for many.

The structure that works is unglamorous and well established. One substantial pillar covering the subject broadly. Around it, focused pieces each answering one specific question thoroughly. Links from cluster to pillar and back, using descriptive anchor text that states the subject rather than “click here”.

Descriptive anchors matter more than they used to, because anchor text is one of the clearest available statements about what a linked page covers, and both crawlers and retrieval systems read it.

There is a threshold effect worth understanding. Partial coverage performs disproportionately badly, because you compete against sites with complete coverage for every fragment and lose most of them. Either cover a subject properly or pick a narrower subject you can cover properly. Half a cluster is the worst position to occupy.

Part nine, platform by platform

The fundamentals transfer across engines. The emphases differ.

ChatGPT

ChatGPT holds the largest share of AI search usage, estimated around 70 per cent. It draws on two distinct routes, live web retrieval and its own training data. That duality matters strategically, because you can be known to ChatGPT through training exposure even on a query where it performs no live search. Long standing presence in widely mirrored sources pays off here more than anywhere else.

Its retrieval leans on Bing’s index, so Bing Webmaster Tools is direct infrastructure for ChatGPT visibility rather than optional housekeeping, and it gets routinely neglected. Wikipedia is disproportionately represented in its citations. Typical answers name three to five sources.

Google AI Overviews and AI Mode

These sit on top of Google’s existing ranking stack, making them the friendliest surface for anyone with an established SEO programme. Content already ranking well is materially more likely to be drawn on. Roughly three to four citations get surfaced.

Two behaviours are specific here. Local relevance is weighted heavily for anything with geographic intent, so Google Business Profile accuracy feeds AI Overviews directly. And because Google controls both the index and the synthesis, structured data has a more plausible route to influence than on platforms that only crawl.

Perplexity

Perplexity is the most citation dense engine, commonly listing five to ten sources, and the most transparent about where answers come from. It weights recency strongly and performs live retrieval on nearly every query.

For most businesses it is the best place to start measuring, for a practical reason. Because it cites generously and shows its working, you can see which sources it favours in your category and reverse engineer the pattern. Treat it as your diagnostic surface even where it is not your largest traffic source.

Gemini, Claude and Copilot

Gemini is the fastest growing platform and wired deeply into Google’s infrastructure, so visibility here tends to follow from Google performance rather than requiring separate work.

Claude synthesises more than it quotes, favouring material with clear logical structure over material that is merely comprehensive. Well argued well ordered writing does better here than keyword dense writing. Integration into Safari will raise its significance considerably.

Copilot runs on Bing’s index and is embedded across Windows and Microsoft 365, giving it substantial reach in business contexts specifically. Where your customers are B2B it deserves more attention than its consumer profile suggests, and the optimisation work overlaps almost entirely with ChatGPT.

Part ten, measurement and the honest position on it

Measurement is the weakest part of this discipline and anyone claiming otherwise is selling something. Most AI search is zero click, so conventional analytics cannot see the majority of your exposure.

Share of voice is the closest thing to a headline metric. Across a fixed basket of prompts, how often are you named? Competitive share tells you who appears beside you and how often, which is frequently more actionable than your own number in isolation.

Citation tracking answers which specific URLs, yours and other people’s, get cited for the questions that matter commercially. Description accuracy is badly under used. When an engine describes your business, is it correct? Being cited inaccurately can be worse than not being cited.

Crawler access can be confirmed definitively from server logs. Referral traffic is small but high intent, and one platform has reported around a tenth of new signups arriving via ChatGPT referrals.

The attribution problem

AI referral traffic is systematically under counted, and understanding why prevents a lot of bad decisions.

Some assistants strip referrer headers, so visits land in analytics as direct. Others send users to a URL the assistant paraphrased rather than linked. A large share of the value never involves a click at all. Somebody reads that you are a good fit, does nothing, and searches your brand name three days later. That converts as branded organic or direct, and nothing connects it to the AI answer that caused it.

Two consequences follow. Rising direct and branded search alongside flat organic may be AI working rather than a plateau. And you cannot build a conventional return on investment case from last click attribution here, so do not try. Measure share of voice and treat branded search volume as the corroborating signal.

Setting up what you can actually track

Build a channel grouping in analytics capturing chatgpt.com, perplexity.ai, gemini.google.com, claude.ai and copilot.microsoft.com. Small numbers, but real and trackable.

Watch branded search volume in Search Console, because a sustained rise in impressions for your brand name is one of the better proxies for AI exposure you cannot otherwise see. Log AI crawler hits so you know which pages the bots favour and how often they return. Record the prompt basket over time in a spreadsheet, which is unglamorous and more reliable than most dashboards.

On tooling

A tracking tool market appeared very quickly, spanning dedicated AI visibility platforms and modules bolted onto established SEO suites. They broadly do the same job. Run prompt sets on a schedule, record whether you are named, track competitors, monitor sentiment.

Two cautions before buying. The category is immature and methodologies differ enough that two tools can report materially different share of voice figures for the same brand, so pick one and stay with it rather than comparing across vendors. And nearly all published statistics in this field come from these same vendors, who have an obvious interest in the numbers being alarming. Establish your manual baseline first. It costs nothing and tells you whether a tool measures anything you cannot see yourself.

Part eleven, when the machine gets you wrong

Almost no guide covers this, and it is the risk most likely to cost you money. Visibility is not automatically good. An AI assistant confidently describing your business incorrectly, to thousands of people, is worse than not being mentioned.

The failure modes are predictable. Outdated pricing or services from a page you changed two years ago. Attribution of a competitor’s weakness to you. Confusion with a similarly named business. Stale leadership or location details. Or the model simply inventing a plausible sounding detail with no source at all.

What to do about it

1. Check deliberately. Ask each major engine what your company does, what it costs and who runs it. Most businesses have never done this once.
2. Trace the source. Engines that cite will show you where the wrong claim came from, and it is usually one specific stale page, sometimes your own.
3. Fix at source. Correct or remove the originating page. Where it belongs to a third party, ask them. Slower, and the only durable fix.
4. Publish the correct version prominently. A clear current well structured page stating the facts plainly gives retrieval something better to find.
5. Re-check on a schedule. Corrections propagate slowly and unevenly, and training data lags live retrieval by a long way.

Add description accuracy to your regular reporting alongside share of voice. Being described correctly is a prerequisite for visibility being worth anything.

A 90 day plan

Every guide lists tactics and almost none sequence them. Order matters, because the cheap diagnostic work frequently reveals that the expensive work is unnecessary.

Days 1 to 14, find out whether you are even visible

1. Audit robots.txt and your CDN for AI crawler blocks, then confirm against server logs.
2. Disable JavaScript and check what survives on your ten most valuable pages.
3. Build a twenty prompt basket and run your manual baseline.
4. Record where competitors appear and which third party pages get cited.

Days 15 to 45, fix the foundations

1. Resolve any crawler or rendering blocks found above. This alone can change everything.
2. Restructure your ten priority pages with one H1, clean hierarchy, answer first openings and self contained passages.
3. Add author bylines with genuine credentials, plus publication and review dates.
4. Implement schema only for content actually on the page.

Days 46 to 90, build authority and measure

1. Write content for the fan out fragments identified earlier.
2. Pursue inclusion in the third party pages already being cited in your category.
3. Publish something genuinely original. Your own data, a real case study, first hand testing. The most durable advantage available, because it cannot be synthesised from other people’s pages.
4. Re-run the prompt basket and compare against baseline.

Mistakes we see most often

• Treating GEO as separate from SEO. They share infrastructure, and splitting them into rival budgets produces two underfunded programmes.
• Mass producing AI content to feed the machines. Fails on Google’s own stated criteria and does nothing for citation.
• Marking up invisible content. A genuine penalty risk rather than a grey area.
• Drawing conclusions from single prompts. The system is probabilistic, so one observation is noise.
• Optimising only your own domain when much of the judgement forms off site.
• Date bumping instead of updating, which Google explicitly names as a red flag.
• Ignoring how you are described. Visibility with an inaccurate description can actively cost you business.

Glossary

Definitions in this field get used loosely. These are the working definitions used throughout this guide.

• Answer engine. Any system responding to a question with a direct answer rather than a list of links, including AI chat assistants, featured snippets and voice assistants.
• Chunking. Splitting a page into passages before retrieval. Retrieval operates on chunks rather than whole documents, which is why self contained sections matter.
• Citation. A named source attached to an AI generated answer. The AI search equivalent of a ranking position, except several get awarded at once and none is fixed.
• Entity. A distinct thing, meaning a company, person, product or place, that a machine can recognise, describe and relate to other things. Entities rather than keywords are what retrieval systems reason about.
• Extractability. How easily a passage can be lifted from a page and remain accurate and comprehensible alone.
• Grounding. Anchoring a generated answer to retrieved sources rather than the model’s own memory. Grounded answers are the ones carrying citations.
• Hallucination. A confidently stated claim with no basis in any source. Relevant here because it can happen about your business.
• llms.txt. A proposed plain text file at a domain root describing site structure for AI systems. Adoption by major engines is unconfirmed.
• Non determinism. The property that identical prompts can produce different answers, and the reason AI visibility is a frequency rather than a rank.
• Query fan out. Decomposition of one question into several shorter sub queries, each searched separately.
• RAG, retrieval augmented generation. Fetching relevant passages and supplying them to the model as context rather than relying on training data alone.
• Share of voice. How often a brand appears across a defined set of prompts. The headline metric of AI search visibility.
• Structured data. Machine readable markup, usually Schema.org, stating explicitly what a page’s content represents.
• YMYL. Your Money or Your Life. Topics affecting health, finances, safety or welfare, where Google applies markedly higher quality standards.
• Zero click search. A search satisfied without any click through to a website. Now the majority of searches.

What nobody can prove yet

A closing section most guides omit, and the most useful thing here. These are open questions, and anyone answering them with certainty is guessing.

• Whether llms.txt influences anything, given adoption is unconfirmed.
• How much weight training data presence carries against live retrieval, and whether that differs by engine.
• Whether the three month recency cliff is a real property of these systems or an artefact of how it was measured.
• The true overlap between Google rank and AI citation, where credible sources contradict each other outright.
• Whether schema markup meaningfully affects LLM citation, as opposed to affecting classical rich results which then affect retrieval.
• How personalisation will change measurement once answers vary systematically by user.

The strategic conclusion follows from that uncertainty. Concentrate effort on things that pay off under every scenario. Pages a machine can actually read. Prose that survives extraction. Genuine expertise made visible. Original material existing nowhere else.

Those hold whether the overlap figure turns out to be 70 per cent or 20 per cent, and whether or not llms.txt ever matters. Everything else is worth testing and worth holding loosely.

Related reading

• generative engine optimisation — the GEO mechanics in isolation
• answer engine optimisation — the AEO side
• how to get cited by LLMs — the practical checklist
• the free crawler checker — confirm the assistants can read your site at all