SEO, AEO and GEO: why the foundation comes first

Three acronyms now compete for the same budget line, and the vocabulary has outrun the substance. It is worth being precise about what each one changes, because the work differs, the measurement differs, and so does the honest expectation of what it can achieve.

Start with the unit being optimised

The cleanest way to separate the three is to ask what thing is being improved. SEO improves a page. AEO improves a passage. GEO improves an entity — the organisation itself, and the consistency of facts about it wherever they appear.

That distinction has practical consequences. A page can rank third and never be quoted in an answer box, because the answer is buried in paragraph nine. An organisation can rank well for every commercial query and still be described wrongly by an assistant, because its own site never states plainly what it does. These are different failures with different fixes.

SEO: can the page be reached, rendered, understood and ranked?

Most underperformance we are asked to investigate turns out to be structural rather than editorial. The content is adequate; the machinery around it is not. The recurring causes are unglamorous and specific.

• Primary content absent from the initial HTML response, arriving only after client-side rendering
• Canonical tags pointing at a different URL than the one that should rank
• Facet and parameter combinations generating thousands of near-duplicate URLs that consume crawl budget
• Pagination or infinite scroll that leaves most of a catalogue unreachable without JavaScript
• A migration that shipped without a redirect map, discarding accumulated authority silently

The measurement discipline matters as much as the fix. Record a baseline — index coverage, impressions, positions, field Core Web Vitals — before making changes. Without it, every subsequent movement is attributed to whatever the person reporting prefers to believe.

AEO: can one specific answer be extracted?

Answer surfaces extract passages, not documents. That single fact determines the work. A page that answers a question implicitly across four paragraphs is far less useful to an extraction model than a page with the question as a heading and a direct forty-to-sixty-word answer immediately beneath it.

1. 01Harvest the questions people actually ask — from support tickets, sales calls, site search logs and search data. Invented questions help nobody.
2. 02Assign each question cluster to exactly one owning page, so two of your own pages stop competing to answer it.
3. 03Restructure: question-shaped heading, direct answer in the opening lines, then the depth a reader wants.
4. 04Add FAQPage, HowTo or QAPage markup only where the visible content genuinely matches the markup.
5. 05Measure question-shaped query performance, and watch whether repeat questions to your support team decline.

GEO: is the organisation described accurately?

Generative systems reconcile what they can find. Where your own site is vague, inconsistent or unreachable, they fill the gap from a stale directory listing or a competitor's comparison page. The result is a description of your business written by your least accurate source — and prospects increasingly arrive quoting it.

The work is closer to entity hygiene than to conventional optimisation. One canonical description of what the organisation is, who it serves and where it operates. The same service taxonomy and terminology everywhere. Machine-readable Organization and Service data that matches the visible page. Same-as links to authoritative profiles. Content structured so a passage can be lifted and attributed: self-contained sections, specific rather than promotional claims, defined terms, visible dates.

Then measure, because GEO without observation is guesswork. Fix a prompt set, replay it across the major assistants on a schedule, and record three things verbatim: whether you are mentioned, whether the description is accurate, and which sources are cited. It is sampling rather than analytics, and it should be reported as such.

If a supplier guarantees a position in an AI answer, they are describing a system none of them controls. Non-determinism is not a marketing problem to be solved; it is a property of the technology.

What all three actually share

The layers differ above the waterline. Below it they are identical, which is why foundation work is the highest-leverage spend available.

• Content present in the initial HTML response — invisible here means invisible to all three at once
• Semantic structure: real heading hierarchy, self-contained sections, no meaning carried only by visual styling
• Structured data that matches what the page shows, rather than what you wish it showed
• Consistent, specific facts about the organisation across owned and third-party properties
• Genuine depth on the subjects you want to be associated with

Note that the last item is not a technical control. Ranking algorithms, extraction models and generative systems all reward content that answers a real question in detail, because that is the property each of them is trying to detect. Depth is the one input that serves every surface simultaneously.

A defensible sequence

1. 01Fix the technical foundation: rendering, crawlability, canonicals, duplication, Core Web Vitals. Everything else is gated on it.
2. 02Establish entity consistency — one description, one taxonomy, structured data that matches the page. It is cheap and it serves all three layers.
3. 03Restructure your highest-value existing content answer-first, rather than publishing more.
4. 04Baseline what assistants currently say about you, so later claims of improvement have something to compare against.
5. 05Only then decide where additional content investment goes, guided by how your buyers actually search.

The order is deliberate: each step makes the next cheaper, and the first two are the ones most often skipped in favour of publishing volume.

Claims worth distrusting

• Guaranteed rankings or guaranteed AI mentions — nobody controls either system
• llms.txt presented as an AI-visibility solution rather than a cheap, unratified convention
• FAQ schema sold as a route to rich snippets in 2026
• Traffic improvements attributed to a change with no recorded pre-change baseline
• GEO reporting with no published prompt set, which makes the numbers uncheckable

The common thread is unfalsifiability. Work that can be checked — index coverage before and after, a field vitals measurement, a prompt set anyone can replay — is worth paying for. Work that cannot be checked is worth being sceptical about, regardless of how current the acronym is.