On this page 13 sections
- Key takeaways
- Who this applies to
- What the terms mean
- What the click data shows
- What is the same
- What is genuinely different
- Where the budget should go
- What to stop doing
- Should you reallocate budget?
- What we sell, and what we tell people not to buy
- When to ignore this entirely
- Frequently asked questions
- Next step
Most of what is sold as generative engine optimisation is search engine optimisation with a new name: be crawlable, be authoritative, answer the question clearly, get referenced by others. Four things are genuinely different. The unit that competes is a passage rather than a page, the outcome is a citation rather than a click, third-party sources carry more weight than your own site, and the measurement is a sampled rate rather than a rank. Budget accordingly: mostly existing practice, with a real but narrow new layer, and a business case that has to survive fewer clicks.
The useful question is not whether GEO is real. It is which parts of it are new, because you are already paying for the rest.
Key takeaways
- The overlap between the two disciplines is large, and that is the honest headline.
- The click data is now measured, not argued. Pew found users click a result on 8% of searches with an AI summary against 15% without one.
- The retrieved unit is a passage, so page-level thinking under-serves you.
- Assistants lean harder on third-party sources than search results pages display. The highest-leverage work is often not on your site.
- There is no rank and no Search Console equivalent. Measurement is sampling, with variance.
Who this applies to
You are being pitched GEO, AEO, LLMO or AI-visibility services, or deciding whether to move budget from conventional SEO, and you want to know what is genuinely different before you pay for a rebrand of what you already buy.
It applies to services businesses, software companies and anyone whose buyers research before contacting. It applies with a different conclusion to publishers monetising pageviews, and that case is called out where it differs.
What the terms mean
Three acronyms are in use for substantially one activity.
| Term | Stands for | What it usually means in a proposal |
|---|---|---|
| SEO | Search engine optimisation | Ranking pages in a list of results, measured by position and clicks |
| GEO | Generative engine optimisation | Being cited or recommended inside a generated answer, in Google AI Overviews and AI Mode, ChatGPT, Perplexity, Claude and Copilot |
| AEO | Answer engine optimisation | The same as GEO, from a vendor who chose the other word |
The proliferation of acronyms is itself a signal about the maturity of the category. Nobody has a settled measurement, so everybody has a name.
What the click data shows
Until 2025 this argument ran on anecdote. It no longer has to.
In July 2025 the Pew Research Center published a study of 68,879 Google searches by 900 US adults who shared their browsing data. When a search returned an AI summary, users clicked a traditional result on 8% of those searches. When it did not, they clicked on 15%. Clicks on a link inside the AI summary itself happened on 1% of visits. Google disputed the methodology; nobody has published a contrary measurement.
Clickstream data from Similarweb, reported in the first half of 2026, put the share of US Google searches that end without any click at about 68%, up from roughly 60% in 2024. The trend predates AI Overviews. They accelerated it.
Two consequences follow, and they point in different directions.
For a publisher, this is a revenue problem, not a marketing channel. Half the clicks on a query is half the pageviews.
For a services business, the picture is different. Your buyer was never going to convert on the click. They were going to read three sources, form a shortlist, and contact one. If the assistant names you on the shortlist, the click was never the point. What you lose is the ability to measure it the way you did.
What is the same
Worth stating plainly, because it is most of it.
Crawlability. If a crawler cannot reach and render your content, nothing else matters. Same requirement, partly different user agents. Check your robots file against the AI crawlers you actually want to admit.
Authority and corroboration. Assistants retrieve from a search index and weight corroborated sources. A domain nobody links to is weak in both systems for the same reason. The Domain Rating that decides your search rankings decides your citation odds too, because the retrieval step is a search.
Clear, well-structured content. Headings that match real questions, direct answers, sensible information architecture. This was good advice in 2015.
Technical hygiene. Status codes, sitemaps, canonical URLs, speed, no duplication.
Genuine expertise. Both systems favour content that says something specific over content that covers a topic generically. The generic version is what the model itself would have written.
If a vendor’s GEO proposal is entirely the above, you are being sold SEO. That may be fine. You might need SEO. The pricing should reflect what it is.
What is genuinely different
1. The passage competes, not the page
Conventional search ranks pages. Retrieval splits documents and pulls passages, so a single strong section can be lifted from an otherwise unremarkable page, and a strong page whose sections all depend on earlier context can fail to produce a usable chunk.
Practically: sections must be self-contained and front-loaded. Each H2 should open with the answer to the question its heading asks, in one or two sentences, before the explanation. It is a real change in how you structure a document, and it is the most substantive item on this list. See writing content that assistants actually cite for the method.
2. The outcome may be no click
Covered above with the data. The design consequence: decide which of the three outcomes you need, because they have different causes and different measurements.
| Outcome | What it means | Who needs it |
|---|---|---|
| Mention | Your name appears in the answer | Brand and shortlist |
| Citation | Your page is the linked source for a claim | Authority, and the only one that can send a click |
| Recommendation | The assistant names you as an option for the buyer’s job | Services and software businesses. This is the commercial one |
A vendor reporting a single “AI visibility score” has collapsed these into one number, and the number will move without telling you what to do about it.
3. Third-party sources carry more weight
A search results page shows a list, and your own page can sit at the top of it. An assistant synthesises across sources, and for vendor-selection questions it leans heavily on roundups, directories, review platforms and comparison articles, because those are the pages that already answer “who are the options”.
The consequence is uncomfortable for anyone selling on-site optimisation: the highest-leverage work is often not on your site at all. It is getting into the sources that get retrieved. That is closer to digital PR and listings management than to technical SEO, and it is the line item most GEO proposals leave out because it is the one they cannot automate.
The practical version: run twenty prompts that a buyer in your category would ask, note every third-party page cited, and count. The five pages that recur are your target list. Getting onto them is the engagement.
4. Measurement is sampling, not ranking
There is no rank. There is no Search Console for assistants. The same prompt produces different answers on different runs, with different citations.
The credible method is a panel of prompts, each run several times, tracking mention, citation and recommendation rates over time, with the run-to-run variance reported alongside the rate. That is a weaker instrument than rank tracking, and it is what the mechanism supports. Any vendor reporting a precise position for a query in an assistant is over-claiming. Ask how it was measured and how many runs the number rests on.
Where the budget should go
| Work | New? | Priority | Why |
|---|---|---|---|
| Crawlability, including AI user agents | Mostly not | First | Binary. Nothing else works without it |
| Answer-first restructuring of existing pages | Partly | High | Helps humans and extraction alike, cheap |
| Third-party presence: directories, reviews, comparisons | Not new, newly weighted | High | Most underrated line, and the one that moves recommendations |
| Passage-level structure on new content | Yes | High | The one genuinely new craft skill |
| Structured data | Not new | Medium | Documented for search, plausible for assistants |
| Prompt-panel measurement | Yes | Medium | Needed to know whether anything moved |
| llms.txt | Yes | Low | Cheap and unproven. See llms.txt and structured data |
| Publishing volume | Not new | Low | Without depth it competes with the model itself |
The pattern: the genuinely new items are structure, measurement and a text file. Two of those are cheap and one is unproven. The high-impact work is mostly re-prioritised existing practice, particularly third-party presence.
What to stop doing
The reallocation conversation usually asks what to add. The more useful half is what to stop, because that is where the budget comes from.
Stop reporting rank as the headline. Position three on a query that now returns an assistant answer above the results is worth less than it was, and reporting it as unchanged success hides a real decline. Report clicks and conversions. Keep rank as a diagnostic.
Stop publishing thin coverage pages. One page per keyword variant was always weak and is now actively counterproductive: retrieval matches passages semantically, so ten near-identical pages compete with each other and none is the clearest answer to anything.
Stop writing the long preamble. Content that spends four paragraphs establishing context before answering was optimised for a dwell-time theory that was never well evidenced. It now also fails extraction. Answer first.
Stop paying for volume without depth. Twenty templated articles a month, briefed from a keyword tool, now compete against a system that produces the same thing instantly. If your content is not carrying something specific, a number, a failure, an opinion with reasoning, its floor has dropped to zero.
Stop treating your own site as the whole surface. For vendor-selection questions, third-party sources carry more weight than your homepage. A quarter of the budget moved from on-site work to review profiles, directories and other people’s comparisons is usually the highest-return reallocation available.
Should you reallocate budget?
It depends on one thing: how your buyers actually find you.
If discovery is search-led, do not reallocate much. Assistants use search indexes; work that improves conventional visibility improves both. Add the structural changes, which are cheap, and start measuring.
If your market has moved early to assistants, technical buyers, developer tools, some professional services, set up the sampled measurement now, because you need a baseline before you can claim anything moved.
If you are a publisher monetising pageviews, this is a strategic problem rather than a channel to optimise. The Pew numbers are the shape of it. The answer is about your business model, not your markup.
If you have no organic visibility at all, neither discipline is your first problem. You need authority and third-party presence before any of this applies, which is the honest answer for most new domains, including this one.
What we sell, and what we tell people not to buy
We sell an AI search visibility engagement, so this section should be read with that in mind.
What is in it: a citation baseline measured across a prompt panel with the variance stated, answer-first restructuring of existing pages, structured data, and identifying which third-party sources get cited in the client’s category. That last item is usually the most valuable and the least technical.
What we tell people not to buy: a GEO retainer for a site with no authority. If a domain has no third-party presence, restructuring its pages optimises something nobody retrieves. The correct sequence is to become findable first, and we say so on the first call, which regularly ends the conversation.
We also decline to report a single visibility score. Mention, citation and recommendation are three different outcomes with different causes, and collapsing them into one number produces a chart that moves without telling anyone what to do. The three-line version is less impressive and more useful.
The uncomfortable part of our own position: this domain is new and has essentially no authority. We are writing about a discipline we are not currently winning at, and the honest framing is that this is an argument from mechanism and from published data, not from our own results. Take the reasoning. Do not take our rankings as evidence either way.
When to ignore this entirely
When your pipeline is outbound or referral. Check where your last ten clients came from before investing in any discovery channel.
When your category has no consensus sources. Assistants improvise where there is nothing good to retrieve, and you cannot fix that from your own site. You may be able to fix it by becoming the source, which is a publishing project, not an optimisation one.
When someone quotes you a guaranteed placement. It does not exist for organic answers. The same prompt gives different citations on different runs.
When you have not measured a baseline. Without one, you will not be able to tell whether anything worked, which suits a vendor and not you.
Frequently asked questions
Is GEO a real discipline or a rebrand?
Roughly three quarters existing practice re-prioritised, one quarter genuinely new: passage-level structure, sampled measurement, and the higher weighting of third-party sources. That is a real difference and a smaller one than the marketing implies.
Will AI search kill SEO traffic?
It is already halving clicks on informational queries where a summary appears, on Pew’s measurement. Queries with commercial intent, where people want to compare and choose, still generate visits, because the buyer needs to see the options. The impact depends heavily on which kind your traffic is.
What is the difference between GEO and AEO?
Different labels for substantially the same activity. Some vendors use AEO for featured snippets and voice answers, and GEO for generated answers. In practice the work overlaps almost completely.
Does GEO replace SEO?
No. Assistants retrieve from search indexes, so a page that cannot rank cannot be retrieved. GEO is a layer on top of search visibility, not a substitute for it. A site with no SEO has nothing for GEO to optimise.
Should we hire a specialist?
Only if they can explain what is different from SEO without hand-waving, and only after you have conventional visibility. Ask them how they measure. The answer tells you most of what you need.
How long before we see anything?
Structural and crawlability work: weeks, if it shows at all. Third-party presence: months. Both need a baseline measured first or you will not see anything at all.
Next step
If you want a baseline before deciding whether any of this deserves budget, that is the right order and it is a small piece of work. The AI search visibility engagement starts with measurement rather than with changes.
Related: Why ChatGPT recommends your competitor · Writing content that assistants actually cite · llms.txt and structured data · AI search visibility