Written and reviewed by the Green Arrow Consultancy team, led by Darren Tyler, founder and chief executive.
They are also not two different jobs, which is the version being sold. The useful answer is more awkward than either: one shared foundation, a genuinely different target at the top of it, and a measurement problem that catches teams out.
SEO optimises for a position in a list of links. GEO optimises for being one of the two or three sources a model chooses to quote. The technical foundation is shared, so most SEO work still counts. What is new is extractable answer blocks, entity resolution, crawler access as its own discipline and measuring citation share. GEO does not replace SEO and should not be a separate budget line.
Start with the part nobody selling a new discipline wants to lead with. The great majority of the technical work is identical, because both surfaces are fed by crawlers that have to fetch, parse and understand a page before anything else can happen.
A page that a crawler cannot reach cannot rank and cannot be cited. A page that renders its content only after a JavaScript execution the crawler does not perform is invisible to both. A site with no internal linking has pages neither surface will discover. A slow origin gets crawled less by everyone. A thin page with nothing specific in it is useless to a ranking algorithm and useless to a retrieval system, for closely related reasons.
Expertise carries over too, and more strongly than it used to. Both surfaces are trying to identify sources worth trusting, and both are increasingly bad to fool. The difference is that a search engine can hedge by returning ten results and letting the user choose, whereas a model committing to two sources in a paragraph has made a much more exposed decision. That raises the bar rather than lowering it.
So if somebody tells you your existing search work is now obsolete, they are either mistaken or selling. The realistic figure, in our experience of running both, is that most of an established technical and content programme continues to earn its keep. What is missing is a specific set of additions, and the additions are cheaper than the replacement being proposed.
Now the part that is real. Three differences matter enough to change how you work.
A search result page is a list, and position four on a list is a meaningfully different outcome from position eleven but still an outcome. An AI answer is a paragraph that names two or three sources. There is no position four. You are in the answer or you are not, and the distribution of outcomes is therefore much harsher: a small number of sources take almost everything and the long tail takes nothing at all. That changes the economics of marginal improvement. Moving from tenth to seventh in a ranking has value. Moving from the eleventh most quotable source to the eighth has none.
SEO optimises a page as a unit, because the unit that ranks is a URL. Retrieval systems chunk your content, embed the chunks and pull back the ones that match. The unit that competes is a passage. This is why a page can rank well and never be cited: it is comprehensive, it is well linked, and every individual paragraph depends on the one before it, so no chunk survives being lifted out. Writing for extraction means making sections self-contained, which is a small change with a large effect and one that traditional SEO briefs never asked for.
This is the difference that actually decides where you spend. Google's AI Overviews are built on top of Google's own index and draw heavily on pages that already rank organically, so strong conventional search performance transfers to that surface fairly directly. ChatGPT assembles its citation set differently and the overlap with organic rankings is considerably lower. Perplexity sits somewhere between and moves. The consequence is that a programme which lifts your Google rankings may show up in AI Overviews within weeks and do very little in ChatGPT, and a team measuring one number across all engines will not be able to see that happening.
Seven dimensions. Read the last two rows together, because they are the pair that catch experienced search teams out.
| Dimension | SEO | GEO |
|---|---|---|
| Unit of success | A ranking position for a URL against a query | Inclusion as one of a small number of named sources inside a generated answer |
| Primary signal | Relevance and authority at page level, links, and behavioural signals in aggregate | Passage-level extractability, entity clarity, and corroboration from independent sources |
| How you measure | Rank tracking, impressions and clicks from search console data, organic sessions | Citation share and mention rate across a fixed prompt set, run repeatedly per engine |
| Typical time to effect | Weeks to months for content, months for authority. Reasonably predictable | Days on live-fetch surfaces, weeks on indexed ones, six months or more for corroboration. Much less predictable |
| Who owns it internally | Marketing, usually with an agency and a working relationship with the web team | The same team, plus a standing line to whoever controls the WAF and CDN, which is normally security or IT |
| What a failure looks like | You are on page three. Visible, diagnosable, and there is a ladder to climb | You are absent. No feedback, no error, no partial credit, and often no idea it is happening |
| Biggest single risk | Losing ground to a competitor investing more | A firewall rule you did not know about making every other investment worthless |
Time-to-effect ranges are from our own engagements and vary widely by sector and by how much independent coverage a business already has.
If a supplier tells you this list is now obsolete, they are describing a market position rather than a technical reality.
Six additions. None of them replace anything on the previous list, which is the point.
A 40 to 70 word self-contained answer near the top of each page, written so it still makes sense with everything around it removed. No SEO brief ever asked for this and it is the highest-yield writing change available.
Making it unambiguous which real-world organisation the page is about, by aligning your structured data, your on-page details and the external records that corroborate them. Ambiguity is a reason for a system to skip a source.
Testing what your edge returns to each AI user agent, on a schedule, and owning the relationship with the team that controls it. This did not previously exist as a marketing responsibility and it is now the highest-risk item on the list.
A fixed prompt set, run repeatedly across engines, recording whether you were named and who else was. Rank tracking does not substitute for it and neither does session data.
Self-contained sections, repeated nouns instead of pronouns, answers before reasoning. Small craft changes that determine whether a chunk of your page survives extraction.
Independent sources stating that you exist and are credible. Adjacent to link building but not the same activity, because an unlinked mention in a source a model trusts can outperform a followed link.
There is a genre of article that opens with a precise percentage for how much AI citations overlap with organic rankings. Treat those numbers with suspicion, including when they support a conclusion you like.
The published figures vary enormously, and they vary for structural reasons rather than because somebody made an error. Different studies use different query sets, and a set weighted towards commercial queries behaves nothing like one weighted towards informational ones. They cover different engines, and the engines genuinely differ. They run at different times, and retrieval behaviour changes without announcement. They define a match differently, since domain-level overlap and URL-level overlap produce very different numbers from the same data. And these systems are non-deterministic, so the same prompt on the same day can return a different source set.
What survives all that variation is a directional finding, and it is worth stating plainly because it is actionable. Google's AI surfaces show high overlap with Google's own organic results, which follows from the architecture. ChatGPT shows substantially lower overlap. Perplexity sits between the two. Anyone quoting a single precise figure for the second of those, without naming the query set and the date, is over-claiming, and the same caution applies to any number we would give you.
The practical instruction from that is simple enough. Do not model your investment on a borrowed percentage. Build your own prompt set, measure your own citation share per engine, and let the numbers you generate decide where the next quarter goes. The citation checklist sets out how we construct that measurement.
The commercial pressure in this market is to sell GEO as a distinct service with a distinct retainer, because a new line item is easier to sell than an addition to an existing one. We understand the incentive and we think it produces bad outcomes for the buyer more often than not.
Here is what actually happens. Two suppliers audit the same website and produce two technical reports with overlapping and occasionally contradictory recommendations. Two content plans target the same pages with different briefs, and the web team is asked to implement both. A structural change requested by one is reversed by the other. Measurement is reported in two formats that cannot be reconciled, so nobody can say which spend caused which movement. And the single highest-value fix, which is usually a firewall rule, sits unowned between them because it belongs to neither content plan.
The version that works is one owner, one audit, one content plan, and the extra capabilities added to the existing team. If your incumbent supplier cannot test crawler access, cannot advise on entity resolution and does not measure citation share, that is a capability gap you can fill with training, a specialist working alongside them, or a change of supplier. It is not a reason to run two programmes against one website.
There is a narrow exception. If AI search visibility is genuinely strategic for you, if a competitor is being cited and you are not, and if your existing supplier is defensive about it rather than curious, a specialist engagement with a defined scope and an end date is reasonable. Scope it as a project with deliverables that hand over, not as a parallel retainer that never ends. That is how our AI search optimisation engagements are structured, and where the answer turns out to be a system rather than a set of pages it moves into AI consulting.
Five steps. Notice that none of them require a second supplier, and three of them are things a competent search programme should arguably already be doing.
One technical audit covering conventional crawlers and AI crawlers together, because the fixes live in the same infrastructure and the same team implements them. Test what your edge returns per user agent and put the result on a schedule. This is where the most common silent failure lives.
Merge your keyword set with the questions buyers actually ask, collected from sales calls, support tickets and the search box on your own site. The combined set is the content brief and the measurement baseline for both surfaces, which stops the two disciplines drifting apart at the planning stage.
Move the answer to the top of your thirty most valuable pages, make the sections self-contained, and add a comparison table where a comparison is being described in prose. This is a fraction of the cost of a content rewrite and it is most of the available gain.
Organization markup with stable identifiers, consistent details across the site and the external records, and the corroborating profiles cleaned up. One job, done once, benefiting both surfaces and the definitions people cite you from.
Rankings and citation share in one document, on one schedule, reviewed by one owner. When a change lifts one and damages the other, you want to see it in the same meeting rather than in two decks six weeks apart.
Including the ones where the honest answer is that nobody knows yet.
No, but it is closer to SEO than the people selling it as a new discipline would like. The technical foundation is shared almost entirely: if a crawler cannot reach and parse your page, neither surface can use it. What genuinely differs is the unit of success. SEO optimises for a position in a list of links. GEO optimises for being one of two or three sources a model chooses to quote. Those are different targets and they reward different writing.
Usually not. Two suppliers working on the same pages with different briefs produces conflicting recommendations, duplicated technical audits and a content calendar nobody owns. The better structure is one team with the extra capabilities added: crawler access testing, extractability review, entity resolution and citation measurement. If your current supplier cannot do those, that is a capability gap to fill rather than a second retainer to sign.
It displaces some of it and changes the shape of the rest. Answers that used to require a click are resolved in the interface, so informational queries lose sessions while the visits that do arrive tend to be further along. The practical implication is that measuring GEO by sessions will make it look like a failure, because the mechanism partly removes sessions by design. Measure citation share, then measure what the surviving visits do.
It varies enormously by engine and by study, and the honest answer is that no single number is defensible. The consistent directional finding across published research is that Google's AI Overviews draw heavily on pages that already rank in the organic top ten, while ChatGPT's citation set overlaps considerably less. The range of figures published for that second overlap is wide enough that quoting one precise percentage is over-claiming, and we do not do it.
Check that AI crawlers can reach you, because a firewall block cancels every other improvement silently. Then rewrite the opening of your most important pages so the answer arrives in the first paragraph rather than the fifth. Those two changes cost very little and are the only two we would make before measuring anything.
The concept does, the practice has shifted. Keyword volume tables are a poor guide to what people ask an assistant, because prompts are longer, more conversational and frequently multi-part. What replaces them is a prompt set: the actual questions your buyers would type, collected from sales calls and support tickets rather than from a volume tool. That set becomes both your content brief and your measurement baseline.
Whoever owns organic search, with a standing line to whoever owns the edge infrastructure. The second relationship is the one that does not exist in most organisations and it is the one the work depends on, because the highest-value fix is usually a firewall rule rather than a content change.
The mechanism has changed even though the activity looks similar. What matters for citation is not the link as a ranking signal, it is the mention as corroboration: an independent source stating that your organisation exists, does this work, and is credible at it. A mention in a trade publication with no link at all can be worth more than a followed link from a directory nobody reads.
The parts that are evidenced are not early: crawler access, clean structure, entity resolution and independent corroboration are all things that were worth doing anyway and are now worth more. The parts that are speculative, which is most of what is being marketed as GEO tactics, are early. Fund the first group and treat the second as an experiment budget with a review date.
Written and reviewed by the Green Arrow Consultancy team, led by Darren Tyler, founder and chief executive.
We will review your current search programme, test what the AI crawlers see, and tell you which additions are worth making and which of the fashionable ones are not.