Services / SEO, AEO & GEO
Three different systems, three different failure modes, one honest caveat.
Classical search ranks documents, answer engines extract passages, and generative engines retrieve then cite. They are not the same problem, and the current research is clear that optimising naively for the last one can damage the first.
The canon
The principles this rests on, named.
Each with who established it, so the claim can be checked rather than taken on our word.
Relevance and position dominate every content tactic.
Critical survey of Generative Engine Optimization, 2026, across 45 studies
A factorial experiment across 252,000 trials found position within the model's context window mattered as much as or more than most content rewrites. Being retrieved is the game.
Quotations, statistics and cited sources measurably raise citation rates.
Aggarwal et al., GEO, KDD 2024 (Princeton, Allen Institute, Georgia Tech, IIT Delhi)
Measured at +41%, +31% and +27% respectively on position-adjusted word count. Keyword stuffing measured at −8%.
Optimising for citation can reduce retrieval.
SAGEO Arena end-to-end pipeline testing, Kim et al. 2026
Rewrites tuned for citation cut top-20 retrieval presence by about 9% and top-10 after reranking by 16%. High citation on a document nobody retrieves is worth nothing.
The tactics do not generalise across domains.
C-SEO Bench, 1,900 queries
Only three of 54 method-domain combinations were significantly positive. Anyone selling a universal GEO checklist is selling something the literature says does not exist.
Answers are unstable; a ranking without a sample size is noise.
Same 2026 survey
Repeated identical queries over 45 days returned Jaccard similarity of 0.34 to 0.42. Cross-engine URL overlap is 0.11 to 0.18, and 53% of domains cited in Google's AI Overviews are absent from the organic top 10.
Retrieval is a speed and structure problem before it is a content problem.
Information retrieval fundamentals; BM25 and its descendants
A page that is slow, badly structured or thinly linked does not get into the candidate set, and nothing downstream can rescue it.
Method
What actually happens, in order.
Crawl, index control, internal link topology, Core Web Vitals, and structure. If the document cannot be found and parsed, content work is premature.
Definitions that stand alone, question-shaped headings, and answers in the first two sentences under them. This is what answer engines lift.
Statistics with sources, direct quotations, and citations. Not because it games anything, but because it is the one intervention with a measured positive effect and it also happens to be honest.
The same name, address, description and identifiers everywhere, plus structured data, so a machine can resolve who you are without guessing.
Multiple engines, three to five paraphrases per query, several time windows, and randomised controls. Anything less cannot distinguish an effect from noise.
On the page, in public. No reviewed technique shows a stable, cross-platform causal effect on organic discoverability. Publishing that is a differentiator, not a weakness.
What goes wrong
The failure modes nobody advertises.
Studied across roughly 300,000 domains with no measurable relationship to citation frequency, and no major AI provider has committed to reading it. Harmless to ship, dishonest to bill for.
Citation rate is easy to measure, retrieval is not, so most work targets the former and quietly degrades the latter.
A single query on a single day, from one location, on one engine. The instability data makes this close to meaningless.
The oldest pattern in local SEO and the most reliably filtered as doorway content. One page with real specifics beats eleven thin ones.
Featured snippets and generative citation have different mechanics. Conflating them means optimising for neither.
What we refuse
A studio is defined as much by this list.
- We do not sell llms.txt as a service line.
- We do not buy links.
- We do not build city-swapped doorway pages.
- We do not report a ranking without its sample size, date range and engine.
- We do not promise a position. Nobody controls the ranking function.
Measured, not felt
How anyone can tell whether it worked.
Retrieval presence, measured before citation, because citation is conditional on it.
Citation share across at least three engines and three paraphrases, with the date window stated.
Core Web Vitals against a throttled mid-range mobile profile, not a desktop lab score.
Entity resolution: the brand returns the correct structured identity to a machine query.