Generative engine optimisation (GEO)

Overview

Generative engine optimisation is the work of becoming the business an AI assistant names when someone asks it for a recommendation. A buyer who once typed "web development company in Lahore" into a search box increasingly asks an assistant to shortlist three and explain the trade-offs. That shortlist is assembled from sources the model can find, parse, verify and attribute, and those are properties of your content you can actually work on.

It helps to be precise about what is happening, because the field attracts a lot of confident nonsense. Models do not rank pages. When an assistant answers a question about suppliers, it is usually retrieving current web sources, reading them, and synthesising an answer with citations. What decides whether you appear is whether your content is retrievable, whether it is specific enough to support a claim, and whether the model can resolve that the entity on your contact page is the same entity mentioned in a directory and in an industry article. That resolution problem is the heart of GEO.

This is why entity consistency does more work here than any single piece of clever content. If your business appears as "ZR Tech" on the website, "ZR Tech Soft" in a directory, "ZR Technologies" on a social profile, and with three different phone numbers across the three, a model has weak evidence that these are one organisation. Every variant fragments the confidence it can place in any claim about you. Making the name, phone, email and description byte-identical everywhere is unglamorous, takes a day, and is the highest-return action available.

The second lever is the shape of your claims. Models cite statements that survive being quoted alone. "We are the leading IT company in Pakistan" is unusable: it is unverifiable, it is what every competitor says, and it carries no information. "ZR Tech delivers 24 services across five practice areas to businesses throughout Pakistan, working remote-first in English and Urdu" is usable: it is specific, checkable, and it answers a question someone actually asked. Writing in the second style throughout a site changes how often that site can be used as a source.

The third lever surprises people. Content that acknowledges where your service is the wrong choice gets cited more than content that does not. When an assistant is asked to compare options, it is looking for sources that discuss trade-offs, because a source that only advocates is a weak basis for a balanced answer. A page that says "choose a template instead if your budget is under this figure" is more likely to be quoted than one that says a custom build is always better. This is a genuinely uncomfortable finding for a marketing function, and it is why the comparison pages on this site name the situations where ZR Tech is not the answer.

The fourth lever is mechanical: give crawlers a clean version of your content. That means allowing the AI crawlers you want in robots.txt, publishing an llms.txt file that maps your site in Markdown, maintaining an llms-full.txt corpus generated from your source data rather than written by hand, and offering a plain Markdown mirror of every page with no navigation, no footer and no styling. None of this is a standard, all of it is cheap, and it removes the noise a model has to strip before it can use you.

In the Visibility Stack model, GEO sits at layer three and depends entirely on the two beneath it. Layer one decides whether a machine can reach and read your pages. Layer two decides whether it can lift a clean answer out of them. Layer three decides whether it will attribute that answer to your business by name. Attempting layer three on a site that is slow, thin or inconsistently named is the most common way this budget gets wasted, which is why an honest GEO engagement starts by checking the two layers below and often recommends fixing those first.

The final thing to say is the uncomfortable one. GEO is the least predictable of the three layers. Model behaviour changes between versions, outputs vary between runs of the same question, and no supplier can guarantee that any assistant will name you. What can be committed to is the work: the entity audit, the consistency fixes, the citable content, the machine-readable files, and a monthly record of what a fixed set of assistants actually says when asked your target questions. Anyone offering more certainty than that is describing something they cannot deliver.

Who this is for

This service earns its cost for the situations below. If none of them describes you, say so on the call and we will point you elsewhere.

• A business whose buyers are technical or research-heavy and have visibly shifted to asking assistants first.
• A supplier in a considered-purchase category where buyers ask for shortlists and comparisons before contacting anyone.
• A company already ranking well in ordinary search but absent when an assistant is asked to recommend suppliers.
• An organisation whose details are inconsistent across directories, profiles and its own site.
• A business willing to publish honest trade-offs, including where a competitor or a cheaper option is the better call.
• Anyone who has completed technical search work and answer structuring and wants the next layer.

What the work covers

How it works

Every engagement runs through these stages. You approve at each one before the next begins.

1. Baseline what assistants say now

   Ask a fixed set of assistants a fixed set of target questions and record the answers verbatim, including who they name and what they cite. This is the only honest starting point, and it is what later reporting is compared against.

2. Audit the entity

   Collect every public appearance of your business name, phone, email, address and description. List the variants. Most businesses are surprised by how many exist, and reconciling them is the first fix because everything else depends on it.

3. Fix consistency at source

   Correct the canonical details on your own site first, then in directories, profiles and listings. The website is the authority; if it disagrees with itself between the footer, the contact page and the schema, nothing downstream can be trusted.

4. Rebuild claims as citable facts

   Rewrite the key pages so their central statements are specific, checkable and self-contained. Add a key-facts block to the about page. Remove superlatives that no source could verify, since they weaken the pages that carry them.

5. Publish comparison and trade-off content

   Build the pages that compare options honestly and name the cases where you are the wrong choice. This is the content type most likely to be cited, and it is the stage where most clients need reassurance before proceeding.

6. Ship the machine layer

   llms.txt, a generated llms-full.txt, ai.txt, Markdown mirrors, crawler permissions and freshness signals. Mechanical, quick, and it is where a site that was invisible to crawlers becomes readable.

7. Re-test monthly

   Re-run the baseline questions against the same assistants and record what changed. Outputs vary between runs, so the useful signal is the trend across months rather than any single answer.

Making the choice

What you get out of it

• Presence in the shortlist stage, which for many categories now happens inside an assistant rather than a results page.
• A consistent entity, which improves how every other channel represents you, including ordinary search and directories.
• Content that is genuinely more useful to human readers, because specificity and honesty read better than superlatives.
• Machine-readable surfaces that also serve aggregators, syndication and any future crawler convention.
• A defensible position, since competitors unwilling to publish trade-offs cannot easily copy the approach.
• Early positioning in a discipline where most Pakistani competitors have done nothing at all.
• A monthly record of what assistants actually say about your category, which is useful market intelligence on its own.

The implementation checklist

Everything below can be done without hiring anyone. It is published because a reader who does it themselves and still needs help is a better client than one who never understood the work.

1. Write down ten questions a buyer would ask an assistant before choosing a supplier like you.
2. Ask those questions in ChatGPT, Claude, Perplexity and Google AI Overviews, and save the answers verbatim.
3. Note which businesses are named and which sources are cited, because those sources are your real competition here.
4. List every public appearance of your business name, phone, email and description, and record every variant.
5. Choose one canonical form of each and correct your own website first, including the footer and the schema.
6. Correct the same details in every directory, profile and listing you control.
7. Add Organization schema with stable identifiers, a knowsAbout topic list and a sameAs array of profiles you actually own.
8. Add a key-facts block to your about page as a clean definition list of checkable statements.
9. Rewrite your five most important claims so each is specific, verifiable and readable without surrounding context.
10. Delete or replace every superlative that no external source could confirm.
11. Publish at least one honest comparison page that names when a competitor or cheaper option is the better choice.
12. Check robots.txt does not accidentally block GPTBot, ClaudeBot, PerplexityBot or Google-Extended.
13. Publish an llms.txt file mapping your key pages in Markdown with one line of description each.
14. Add a visible last-updated date and a dateModified value in structured data on every content page.
15. Repeat the assistant questions monthly and keep the answers, because the trend is the measurement.

How to measure it

The primary measure is a citation log. Ask the same questions to the same assistants each month and record whether you were named, whether you were cited as a source, and who was named instead. It is imperfect because outputs vary between runs, and it is still the only measure that reflects the actual outcome you are buying.

The secondary measure is referral traffic from assistant domains in your analytics. Volumes are small compared with search and the attribution is incomplete, but the trend is informative, and these visitors typically arrive further through their decision than search visitors do.

A third measure worth keeping is entity health: how many variants of your name, phone and description still exist across the web, tracked as a number that should fall to one. It is fully within your control, unlike the others, which makes it the most honest thing to hold a supplier accountable for.

What is not a measure is a supplier screenshot of one favourable assistant answer. Any question can be phrased to produce a flattering response, and a single run is not evidence. Insist on a fixed question list, a fixed set of assistants, and a record kept over months.

Common mistakes

Blocking the crawlers you want to be read by

A surprising number of sites disallow AI crawlers through a blanket robots.txt rule copied from somewhere else, then invest in content hoping to be cited. Check the file explicitly for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended. This takes two minutes and invalidates everything else if it is wrong.

Publishing claims no source could verify

Superlatives such as "leading", "best" and "number one" are unusable to a model because they cannot be checked and every competitor makes the same claim. Worse, they reduce confidence in the surrounding text. Replace each with a specific, checkable statement, or delete it.

Letting the entity fragment across sources

Different names, numbers or descriptions in different places split how a model represents you, so evidence about your business does not accumulate. Pick one canonical form and enforce it everywhere, including the schema, the footer, every directory and every social profile.

Writing only advocacy

A page that argues for your service without acknowledging when it is the wrong choice is a poor source for a balanced answer, and assistants are assembling balanced answers. Publishing the cases where you are not the right supplier is what earns the citation in the cases where you are.

Treating llms.txt as the whole discipline

The file helps and it is cheap, but it is a site map, not a strategy. A clean llms.txt pointing at vague, inconsistent, unverifiable content simply helps a model find content it cannot use. The content work is the substance; the file is the delivery mechanism.

Expecting the predictability of search work

Model outputs vary between versions and between runs of the same prompt. A month where you are named less is not necessarily a failure, and one where you are named more is not necessarily a win. Judge the trend across several months, and be sceptical of any supplier presenting a single screenshot as proof.

What this is not

GEO is not a way to control what a model says about you. Nobody can guarantee an assistant will name a business, and any supplier offering that guarantee is either misunderstanding the mechanism or misrepresenting it. What is controllable is the quality, consistency and machine-readability of the evidence a model finds.

It is not a replacement for search work. The overwhelming majority of commercial intent still lands in ordinary search results, and assistants themselves frequently retrieve from the same web index. A business that abandons SEO for GEO will lose more than it gains, which is why this is layer three rather than a substitute for layer one.

It is not reputation management. If assistants say something inaccurate about your business because inaccurate information exists publicly, the fix is correcting the sources, which is a different service with a different process. GEO improves how well a model can represent you; it does not suppress what already exists.

And it is not worth buying first. If your site is slow, thinly indexed or inconsistently named, that work comes first and costs less. An honest GEO conversation frequently ends with a recommendation to spend the budget on technical search work instead, and that is the recommendation you will get here if it applies.

This will not work if…

• Your site is not indexed or is blocked from the crawlers that assistants retrieve from.
• Nobody in your category asks assistants for recommendations, which is still true in some local trades.
• You are unwilling to publish specific figures, methods or trade-offs about how you work.
• Your business details are controlled by third parties who will not correct them.
• You need measurable enquiry growth within one quarter, since GEO is the slowest and least predictable layer.
• You expect a guarantee, which no honest supplier in this discipline can give.

Tools and platforms

What the work is actually built with. Named so you can verify it and take it elsewhere if you ever want to.

What you physically receive

• A baseline record of what named assistants currently say about your category
• An entity audit listing every variant of your name, phone, email and description
• A connected schema graph with stable identifiers, knowsAbout and sameAs
• A key-facts block on your about page written as checkable statements
• At least one honest comparison page naming when you are the wrong choice
• llms.txt, a generated llms-full.txt, ai.txt and Markdown page mirrors
• Corrected robots.txt crawler permissions
• A monthly citation log kept against a fixed question list

Realistic timelines

Ranges, not promises. The figure in your scope document is the one that counts.

Questions about generative engine optimisation

Last updated 8 September 2026