Generative Engine Optimization Agency

Updated September 2026 · Written and maintained by the Progression Agency strategy team

We do this work and we are going to argue against most of how it is sold. Generative engine optimization is roughly 80% SEO fundamentals with the emphasis moved and 20% genuinely new — and the new part costs a few hundred dollars once, not several thousand a month forever. This page separates the two precisely, explains what actually gets a page cited by AI answers, shows which queries are really losing clicks and which are barely affected, and gives you the questions that expose whether somebody selling GEO understands it.

What is GEO? The short definition

Answer first: GEO stands for generative engine optimization, and it is the practice of making a site likely to be quoted or cited by AI-generated answers rather than merely ranked by a search engine. When people ask what is GEO in a marketing context, that is the whole of it.

Why the GEO term is confusing

The same three letters already meant geographic targeting in advertising long before generative engines existed, so a request to improve GEO can mean two entirely different pieces of work. It is worth confirming which one somebody means before quoting for it.

GEO, AEO and SEO

SEO aims at a ranked list of links. AEO, answer engine optimization, aims at being the answer shown above that list. GEO aims at being cited inside an answer a model generates. In practice the work overlaps heavily, which is the honest version most vendors leave out.

What GEO actually is

Generative engine optimization is the practice of making a site likely to be cited by AI-generated answers — AI Overviews in Google, and assistants like ChatGPT, Claude, Perplexity and Gemini. The underlying shift is real. The service being sold around it is mostly not new.

The three things that are genuinely new

A deliberate robots policy for AI crawlers. A plain, machine-readable statement of what your business is and does. And monitoring whether you actually appear in AI answers. That is the honest list. All three together are a few hours of work and perhaps $400 of setup.

Everything else was already best practice

Structured data, entity clarity, original research, clean heading hierarchy, earning citations from real publications, crawlability. Every one of these was correct advice five years ago and is being resold now at a premium because the label changed.

What actually gets you cited

Answer first, context second

This is the single most consequential writing change. AI systems extract passages, so a page that states its answer in the first two sentences of a section gets quoted and a page that builds to the answer over six paragraphs does not — regardless of which answer is better. Every section on this page is written that way deliberately.

Be specific enough to be quotable

Vague, hedged prose is unquotable because there is nothing to lift out of it. Numbers, dates, named sources and concrete claims give a system something to cite. ‘Prices vary depending on your needs’ will never appear in an answer; ‘$1,500 to $2,500 a month for a small local business’ might.

Publish something that does not already exist

These systems have already ingested every rewritten summary of every common topic. What they do not have is your data — your audit findings, your pricing, your benchmarks, your worked examples. Original material is the only durable advantage in this, and it is the hardest to fake.

Write sections that stand alone

Because that is exactly what happens to them. A section that only makes sense after reading the previous three will be extracted incoherently or not at all. Self-contained sections are better for human readers too, which is usually how you can tell a genuine practice from a trick.

Which queries are actually affected

This is the part most GEO marketing leaves out. The impact is wildly uneven, and for a great many businesses — particularly local service businesses — it is close to negligible.

Where clicks are genuinely being lost

Definitional queries, factual lookups and simple how-to content. If somebody asks what a term means, the answer appears and the click frequently does not follow. If your traffic was built on explainer content, you have felt this.

Where almost nothing has changed

Local service search, transactional queries and brand navigation. Somebody looking for a plumber near them still gets a map pack and still calls somebody. Somebody buying a product still has to reach a store. The map pack and the checkout are not going anywhere.

What we tell clients to do about it

Shift content budget toward commercial and comparison content, which loses far fewer clicks, while keeping definitional content for the authority it builds rather than the traffic it used to bring. That is a reallocation, not a new service line.

What GEO costs, and what it should cost

This is the argument in one chart. Agencies are pricing GEO at $1,500 to $3,500 a month as a separate line, for work that overlaps almost entirely with an SEO retainer you are probably already paying for. The genuinely additive work is a few hours.

Why we include it rather than charge for it

Because it is SEO. Writing clearer answers, implementing structured data properly, building entity clarity and producing original data are things a competent SEO engagement should already be doing. Charging separately would mean either we were not doing them before, or we are billing twice.

When a separate engagement is legitimate

One case: a one-off project to make an existing site AI-legible when you are not running an SEO retainer at all. Crawler policy, entity statement, structured data audit and a monitoring baseline. That is a defined project worth a few thousand dollars once, and we will scope it that way.

What is not legitimate

A recurring monthly GEO retainer alongside a recurring monthly SEO retainer, delivering overlapping work. If somebody proposes both, ask them to list what appears on the GEO invoice that does not appear on the SEO one. The answer is usually short.

The AI crawler decision

The one genuinely new technical decision, and it deserves a real answer rather than a default.

What the choice actually is

AI companies operate crawlers separate from search crawlers — GPTBot, ClaudeBot, PerplexityBot, Google-Extended and others. You can allow or disallow each in robots.txt. Blocking them removes you from the answers those systems generate. Allowing them means your content is used to produce answers that may not send a click.

What we recommend, and why

Allow them, for most businesses. Being cited builds awareness and branded search even without a click, and being absent from AI answers entirely is a worse position than being cited without attribution traffic. Publishers with a subscription model face a genuinely different calculation and sometimes reach the opposite conclusion legitimately.

Google-Extended specifically

Blocking Google-Extended affects Gemini and related AI products without removing you from Google Search itself. That separation matters, and agencies that do not know it exists will either block too much or nothing at all.

The honest caveat

Crawler compliance is voluntary. A robots directive is a request, and not every crawler on the internet honors it. Treat this as directional control rather than a guarantee.

Entity clarity, and why it matters more now

Being unambiguously identifiable as a specific business is worth more than keyword repetition ever was, and it is a solvable problem rather than a competitive one.

What entity clarity means practically

Your business name, address, description and services stated identically everywhere they appear — your site, your Google Business Profile, your citations, your social profiles — and confirmed with Organization or LocalBusiness structured data so a machine does not have to infer it.

Why inconsistency costs you

If a system cannot confidently determine which business you are, it will not cite you as a source for anything. Three slightly different business names across the web is a genuine problem in a world where the answer is being assembled rather than listed.

The about page nobody optimizes

A clear, factual page stating who runs the business, where it operates, what it does and how long it has done it. Rarely ranks for anything, disproportionately useful for establishing what you are. Most sites have a page of adjectives there instead.

How to measure any of this

Honest answer: manually, partially, and better than it was a year ago. There is no clean reporting API for AI answer citation, and any agency presenting a precise GEO dashboard is presenting an estimate dressed as a measurement.

What we actually do

Define a set of prompts that matter for your business, check them across the major assistants on a fixed schedule, and record where you appear and where competitors do. It is manual, it is a sample rather than a census, and we say so.

The Search Console signature

Impressions holding steady while clicks decline on informational queries is the classic pattern of answers absorbing clicks. It looks like an SEO problem and frequently is not. Recognizing it prevents a lot of wasted remediation.

Referral traffic from AI tools

Small but real and growing, identifiable in analytics by referrer. Worth segmenting so you can see it separately rather than having it absorbed into direct traffic.

Branded search as the honest proxy

If people encounter you in an AI answer and later search your name, that appears in branded search volume. It is indirect, it lags, and it is currently among the more reliable signals available.

The crawlers, one by one

Each AI system runs its own crawler with its own directive. Treating them as a single decision is the most common mistake, because they do genuinely different things.

The distinction that matters most

Google-Extended is separate from Googlebot. Blocking Google-Extended keeps you in Google Search while removing you from Gemini and related AI surfaces. Agencies unaware of this either block nothing or block Googlebot by accident, and the second is catastrophic.

How to actually set it

The directives go in robots.txt at your domain root, one user-agent block per crawler, and take effect the next time each bot fetches the file. Test the result rather than assuming it — a misplaced wildcard that accidentally disallows Googlebot is a genuinely catastrophic and genuinely common mistake, and it is invisible until traffic disappears. Google publishes a robots.txt tester and it takes two minutes to confirm the file does what you intended rather than what you typed. Do it once, then check again after any hosting or platform change, because those quietly overwrite the file more often than anyone expects.

Why we default to allowing

• Being cited builds awareness and branded search even when no click follows.
• Absence from AI answers is a worse competitive position than presence without attribution.
• The traffic these systems do send converts unusually well, because the user arrived pre-qualified.
• Blocking does not prevent competitors being cited in your place — it guarantees it.
• Compliance is voluntary anyway, so blocking is partial control rather than real control.
• The decision is reversible in minutes, unlike most SEO decisions.

Who should genuinely consider blocking

• Subscription publishers whose entire product is the content itself.
• Businesses whose proprietary data has direct licensing value.
• Sites under active licensing negotiation with an AI company.
• Anyone whose content is legally restricted from redistribution.
• Nobody else, in our experience.

Writing for extraction, in practice

The single highest-return change available, and it costs nothing but discipline. Here is what it actually looks like at the sentence level.

The structure that gets quoted

1. State the answer. First sentence of the section, plainly, with no preamble.
2. Qualify it immediately. The condition or exception, in the next sentence.
3. Give the specific. A number, a date, a named source, a concrete example.
4. Explain the reasoning. Now, once the answer is already on the page.
5. Handle the edge case. Where the answer does not hold, and why.
6. Move on. One section, one question, self-contained.

What this replaces

The conventional structure of building context first and arriving at the answer last. That reads beautifully in an essay and fails completely when a machine is extracting a passage, because the passage it lifts is the one containing the answer — and if that is buried in paragraph six, it lifts somebody else’s paragraph one instead.

Headings as questions

Section headings phrased as the question a person would actually ask work better than clever ones, because the match between question and answer is what a retrieval system is looking for. ‘How much does GEO cost’ beats ‘Investment considerations’ every time.

Tables and lists are extracted more readily

Structured content is easier to lift cleanly than prose. A comparison expressed as a table is more likely to be used than the same comparison written as three paragraphs, which is one reason this page contains as many as it does.

What not to do

• Do not front-load a section with your brand or a call to action.
• Do not spread one answer across several sections.
• Do not hedge everything into unquotability.
• Do not use headings that make sense only in sequence.
• Do not bury numbers inside long sentences.
• Do not write anything you would not be happy to see quoted with your name on it.

Original data, and how to produce it without a research budget

The most durable advantage in this and the piece most businesses assume they cannot do. You almost certainly have publishable data already.

What counts as original data

• Aggregate findings from work you have already done — audit results, common failure rates, typical outcomes.
• Your own pricing, published plainly, which most competitors refuse to do.
• Benchmarks from your own client base, anonymised and aggregated.
• A survey of your own customers, which costs almost nothing to run.
• Timings and costs from projects you have completed.
• Comparisons you had to make internally anyway before choosing a tool or approach.

Why it works

AI systems have already ingested every rewritten summary of every common topic. They do not contain your numbers. When a system needs a figure to support an answer, it cites whoever published one — and there are far fewer of those than there are people restating general advice.

The bar is lower than you think

You do not need a formal study. ‘Across the 40 site audits we ran last year, 71% had unoptimised images’ is publishable, useful, citable and honest. State the sample size and the method, do not overclaim, and it becomes a reference point.

What to avoid

Inventing plausible-sounding statistics, which is both dishonest and eventually traceable. Citing your own unsourced numbers as if they were industry figures. And presenting a sample of twelve as if it were representative of everything.

Reading the impact on your own site

Before spending anything on GEO, establish whether AI answers are actually affecting you. For a large share of businesses the honest answer is barely.

The diagnostic order

1. Segment Search Console by page type before drawing any conclusion.
2. Compare informational pages against commercial pages separately.
3. Check whether impressions held while clicks fell, or whether both moved.
4. Rule out ordinary technical causes — indexation, canonicals, a bad deploy.
5. Only then attribute anything to AI answers.
6. Record a baseline so the next comparison means something.

Why this order matters

AI answers have become a convenient explanation for any traffic decline, including declines caused by perfectly ordinary technical problems that are fixable. Attributing a drop to AI when it was a canonical error means the canonical error never gets fixed.

Structured data, done completely

The most under-implemented item on this list. Most sites carry partial structured data shipped by their theme and nobody has checked whether it is complete or valid.

Partial is the normal state

A theme typically emits Organization and BreadcrumbList and stops. Service, FAQPage and DefinedTermSet are rarely present, and those are precisely the types that tell a machine what a page is for. Completing them is an afternoon on most sites.

Validate, do not assume

Structured data fails silently. It can be present, malformed and doing nothing, and nothing in the site will tell you. The Schema.org validator and Google’s Rich Results Test both check it free, and running them is not something to take on trust from an agency.

What it does and does not do

It does not make you rank. It makes you legible — a machine can confirm what a page is rather than inferring it from prose, which makes citation more likely and enables rich results. Anybody selling schema implementation as a ranking service is overselling genuinely worthwhile hygiene.

A realistic ninety-day plan

If you want to act on this without hiring anybody, here is the order that produces the most for the least.

What that costs

Roughly a week of work spread across three months, and no money. It is most of what a GEO retainer would deliver in the same period, which is the uncomfortable point of this page.

What it does not cover

Sustained original content production, ongoing technical maintenance and competitive content strategy — which is what an SEO retainer is actually for, and where hiring somebody genuinely earns its keep.

Further reading, grouped by what you are trying to do

Everything below is written to be useful on its own, and none of it is gated.

What we got wrong about this, and changed

Worth stating, because an agency that has never revised its position on a fast-moving topic has probably not been paying attention.

We initially recommended blocking AI crawlers

When GPTBot first became blockable, our default advice was to block it — the reasoning being that content used without a click was value extracted for nothing. We were wrong. What actually happened is that businesses which blocked disappeared from AI answers while competitors were cited in their place, and the awareness cost exceeded the extraction cost. We changed the recommendation and told the clients who had followed the earlier one.

We over-weighted the traffic threat

The early modeling assumed broad click loss across query types. The reality is severely uneven — definitional and factual queries lost heavily, commercial and local queries barely moved. For most of our client base, which is local service businesses, the practical impact so far has been small. Saying otherwise would have sold more work.

We under-weighted original data

We treated it as one tactic among several. It has turned out to be the only durable differentiator, because everything derivative is already inside these systems. It is now the first thing we look for when a client asks how to become citable.

What we still do not know

How these systems weight sources, how consistent citation is over time, whether current patterns survive the next model generation, and how much of what appears to work is correlation. Anybody claiming certainty here is claiming access to information that is not public.

Questions we get asked on calls about this

‘Our traffic dropped — is it AI?’

Usually not, or not only. Segment by page type first. If informational pages lost clicks while impressions held and commercial pages are unchanged, that is the AI signature. If everything dropped together, it is an ordinary ranking or technical problem and blaming AI means it never gets fixed.

‘Should we stop investing in SEO?’

No, and the logic is backwards. The work that makes you citable by AI systems is largely the work that makes you rank — good content, clean structure, structured data, earned authority. Reducing SEO investment to fund GEO would mean spending less on the thing that produces both.

‘Can we just add a page about AI to rank for it?’

No. Writing about AI does not make you citable; being genuinely useful and specific on your own subject does. This page exists because we have an opinion about how GEO is sold, not as an attempt to rank for a trend.

‘Our competitor appears in ChatGPT and we do not — why?’

Most commonly: they have published something specific and quotable and you have published general advice. Occasionally: your entity is ambiguous and the system cannot confirm who you are. Rarely: you have blocked a crawler and forgotten. Check in that order.

‘How long until this settles down?’

It will not, in any timeframe worth planning around. The sensible response is to do the things that work regardless of which system is asking — clear answers, verifiable claims, original material, unambiguous identity — rather than optimizing for a specific model’s current behavior.

How AI answers actually assemble a response

Understanding the mechanism makes the tactics obvious, and it removes most of the mystique the category is sold on.

Retrieval, then generation

These systems do not recall your page from training. For anything current they retrieve candidate documents at query time, then generate an answer from what they retrieved. That is why crawlability and freshness still matter, and why being in the index of whatever the system retrieves from is the precondition for everything else.

Passages, not pages

Retrieval operates on chunks. A long page is split, and individual passages compete on their own. This is the mechanical reason self-contained sections outperform ones that depend on preceding context, and why a strong section on a weak page can be cited while a weak section on a strong page is not.

Corroboration across sources

Systems prefer claims they can find in more than one place, and prefer specific claims they can attribute. A number stated once by you and nowhere else is citable but uncorroborated; a number consistent with other published figures is both. This is an argument for accuracy rather than novelty for its own sake.

Recency weighting varies by query

For anything with a temporal element — pricing, platform features, best practice — recent material is strongly preferred. This makes content maintenance a citation strategy rather than housekeeping, and it is why dated pages with stale figures quietly stop being used.

Why this makes the tactics unsurprising

Answer plainly so the passage is self-contained. Be specific so there is something to attribute. Be accurate so you corroborate rather than contradict. Stay current so recency weighting works for you. Be crawlable so you are in the candidate set at all. Every recommendation on this page falls out of the mechanism.

Where this is likely to go

Forecasting is cheap and usually wrong, so this is framed as what we are planning around rather than what we predict.

Attribution will probably improve

Citation display has already become more prominent across several systems, driven by publisher pressure and by users wanting to verify. We plan on the assumption that being the cited source becomes more valuable rather than less, which favors original data over volume.

Measurement will get better but stay partial

Some reporting has started appearing in Search Console and analytics tools. We do not expect anything resembling rank tracking, because there is no stable ranking to track — answers vary by phrasing, session and model version.

The fundamentals will not move

Crawlability, clear structure, accurate specific content and earned authority have survived every search change of the last fifteen years. They are the parts of this we would still be doing if every AI product disappeared tomorrow, which is a reasonable test of whether advice is durable.

What we would revisit

The crawler recommendation, if attribution traffic materially improves or if licensing markets develop. It is a reversible decision and we treat it as one rather than a principle.

GEO for specific business types

The impact and the sensible response vary enormously by what you sell, and generic advice here is worse than useless because it leads people to spend on something that will not affect them.

Local service businesses

Barely affected so far, and the correct response is mostly to do nothing beyond the free items. The map pack still dominates local intent, people still need to contact somebody, and a plumber does not lose a job because an AI explained what a burst pipe is. Set the crawler policy, fix entity clarity, and spend the rest on local SEO.

Ecommerce

Product discovery is beginning to move, and comparison queries are where it will land first. Product structured data becomes more valuable, not less, and original material — genuine buying guidance, real sizing and fit information, honest comparisons — is what gets a store cited rather than a review site.

B2B and professional services

The most affected of the commercial categories, because so much B2B content is explanatory. If your pipeline depends on educational content that answers questions, some of that traffic is going. The response is to move toward comparison, pricing and genuinely proprietary material rather than to produce more explainers.

Publishers and media

The hardest position, and the one where blocking is genuinely arguable. If the content is the product, being summarized without a click is a direct revenue loss rather than an awareness gain. This is the one category where we would not push back on a decision to block.

SaaS

Documentation and how-to content is heavily absorbed; comparison and alternatives pages are not. Integration pages, migration guides and honest competitor comparisons hold up well, and those were always the higher-intent pages anyway.

Healthcare, legal and financial

These systems are cautious with regulated topics and lean toward authoritative, clearly attributed sources. Named credentialed authors, cited primary sources and unambiguous organizational identity matter more here than in any other category — which happens to be exactly what these industries should have been doing anyway.

A note on how this page is written

This page argues against the standard pricing of a service we sell, which is unusual enough to be worth explaining.

We wrote it in the format described in it: every section answers its question in the first sentence or two, claims carry numbers where numbers exist, sections stand alone, and headings are phrased as questions people actually ask. If the argument here is right, this page should be citable. If it is not, that is a useful data point too.

The percentages in the click-loss chart are our own estimates from client data rather than published research, and we have labeled them that way rather than presenting them as industry figures. The crawler defaults are our recommendations, not consensus. The pricing comparison reflects quotes we have seen clients receive, which is a sample rather than a survey.

We would rather be specific and occasionally wrong than vague and unfalsifiable. If any of this stops being true — and some of it will — the page gets updated rather than quietly left up, and the change gets noted rather than hidden. There is a section above listing what we already got wrong about this and changed, which is the same commitment applied retrospectively rather than as a promise about the future.

How to judge anybody selling you GEO

The question that settles it

‘What is on the GEO invoice that is not on the SEO invoice?’ A good answer is short, specific and honest — crawler policy, entity statement, monitoring. A bad answer is a list of things that were already SEO, delivered confidently.

Claims that should end the conversation

A proven formula for AI rankings. Guaranteed appearance in AI Overviews. A precise percentage improvement in AI visibility. Nobody has any of these, because the systems are opaque, change frequently and do not publish ranking criteria.

What a good provider sounds like

Willing to say which parts are new and which are relabelled, honest that measurement is partial, focused on original content and structured data rather than tricks, and disinclined to charge you twice for the same work.

What we actually do

Inside every SEO retainer

Content written to answer plainly and be extractable. Structured data implemented completely rather than partially. Entity clarity across the site and every profile. Original data wherever the subject supports it. An AI crawler policy set deliberately. And a prompt set monitored on a fixed schedule.

As a standalone project, if you have no retainer

An AI-legibility audit and implementation: crawler policy, machine-readable business statement, structured data audit and correction, entity consistency check across profiles, and a monitoring baseline you can track yourself afterwards. Scoped as a project with an end, not a subscription.

What we will not sell you

A recurring GEO retainer alongside an SEO retainer. If we are already doing your SEO, this work is included, and telling you otherwise would be charging twice for the same activity.

What changed, in order

Why definitional content still matters

It earns fewer clicks and it still builds the topical authority that makes commercial pages rankable. Deleting your explainer content because it lost traffic is a common overreaction that removes the foundation under the pages that do convert.

Why comparison content became more valuable

Because AI answers are worse at it. Comparing options genuinely requires seeing the options, and people still click through to evaluate. Comparison, alternatives and pricing content lose far fewer clicks than definitional content, which is why we now weight budget toward them.

Why original data became the moat

Everything derivative is already inside these systems. The only thing they cannot have is what you have not published yet — your numbers, your findings, your worked examples. This is also the hardest advice to act on, which is precisely why it is durable.

Generative engine optimization services: what a real scope contains

Answer first: generative engine optimization services should cover crawler access, entity and structured-data clarity, answer-first content restructuring, original data production, citation monitoring and first-party measurement. If a scope has no crawler access work and no measurement, it is content writing with a new label on it.

The conclusion: everything in that table is verifiable by you without the agency’s help. That is deliberate, and it is the fastest way to tell a genuine generative engine optimization service from a rebranded content retainer.

Generative engine optimization companies: how to tell them apart

Answer first: ask what they would change on your robots.txt, how they would measure a citation, and what original data they would help you publish. Firms doing real work answer all three immediately; firms rebranding content work answer none of them.

Search engine optimization providers and GEO: the same suppliers, a wider brief

Answer first: most search engine optimization providers now sell GEO as an extension of what they already do, and for the technical half that is legitimate — crawler access, structured data and entity clarity are all SEO disciplines. The genuinely new parts are answer-first restructuring, original data production and citation measurement, and those are the parts to interrogate.

The AEO, GEO, AI SEO naming problem

Three acronyms for substantially the same thing, which is itself a signal about how the category is being sold.

What each supposedly means

Answer Engine Optimization, Generative Engine Optimization and AI SEO are used more or less interchangeably. Where distinctions are drawn, AEO tends to mean optimizing for direct answers generally, GEO for generative AI systems specifically, and AI SEO is a catch-all. Nobody applies them consistently.

Why the proliferation matters

New acronyms create the impression of new disciplines requiring new budgets. When an industry generates three names for one practice within eighteen months, it is usually the marketing moving faster than the practice.

What to do about it

Ignore the label and ask what the work is. If somebody cannot describe the specific activities in plain language without the acronym, there is nothing underneath it.

What “leading GEO services” should mean in a field this young

Any claim to provide leading generative engine optimization services in the AI industry deserves scrutiny, because the discipline is roughly two years old and nobody has a decade of results to point at. What can honestly distinguish a provider at this stage is narrow and checkable: whether they measure citation and inclusion in AI answers systematically rather than anecdotally, whether they can show how a specific page’s treatment changed across model updates, and whether they are candid that a large part of what works is the same structural clarity, source quality and factual specificity that has always served search. Providers claiming proprietary methods that guarantee inclusion in generated answers are describing something the systems do not currently expose. The reasonable posture in a young field is to say what is measured, what is inferred and what is unknown — and to price accordingly.

Frequently asked questions