Which AEO tactics actually work?
Last updated
I sell an Answer Audit, so I have an obvious interest in this category being real. That is exactly why the grades are attached to studies you can open rather than to my opinion, and why four of the nine say do not bother.
Nine tactics, graded on what the research shows.
| Tactic | What the evidence says | Grade |
|---|---|---|
| Keep pages updated, with a visible date | 75% of AI-cited pages had been updated within twelve months, but only 42% were originally published in that window: freshness is manufactured by updating, not by publishing more. Seer, 7,683 cited pages, 2026. Ahrefs found AI-cited content 25.7% fresher than organic results across 17M URLs. | Do |
| Answer in the first third of the page | 44% of verified ChatGPT citations point at text in the first 30% of the page; 31% the middle, 25% the last third. Indig, 18,012 citations, 2026. Google’s own guidance adds that there is no requirement to break content into tiny pieces. | Do |
| Question-shaped pages that match a sub-query | Engines rewrite one question into several searches you never see — Google documents this as query fan-out. A page whose heading is the buyer’s question, answered in paragraph one, is the shape retrieval selects. Observational, and consistent with the passage data above. | Do |
| Consistent entity: same name, role and description everywhere | Branded web mentions correlated with AI visibility at r = 0.66 to 0.71 across 75,000 brands; Domain Rating at 0.27 to 0.33 and backlinks barely at all. Ahrefs, 2025. Correlation, not causation — but consistency is the precondition for a mention to be attributed to you at all. | Do |
| Concrete facts: prices, timelines, definitions | Sold as a 34% lift from adding statistics. That number comes from a paper measured on a synthetic engine; the NeurIPS 2025 replication on live models found statistics at −0.80 to 0.00 rank change and quotations at −0.19 to 0.06. Keep the specifics because they make you quotable and because a buyer needs them, not because of the 34%. | Cheap, unproven |
| Publishing an llms.txt file | Across the server logs of 137,210 domains, 97% of published llms.txt files received zero requests, and 1.1% of the requests that did arrive came from retrieval bots. Ahrefs, 2026, corroborated independently. Google: you do not need to create new machine-readable files. Ten minutes to publish, so harmless — but nobody should invoice for it. | Cheap, unproven |
| Schema markup, sold as an AI lever | JSON-LD added to 1,885 pages and tracked against matched controls: AI Overviews −4.6%, AI Mode +2.4%, ChatGPT +2.2%. No meaningful movement. Ahrefs, 2026. Google’s AI guidance: there is no special schema.org markup you need to add. Schema still earns its keep for rich results in ordinary search — sell it as that. | Skip as AEO |
| “GEO lifts visibility by 40%” | The headline on most agency pages. It comes from a 2023 paper run on a self-built synthetic engine with the document already in the model’s context window, so discoverability was never tested. The 2025 replication found the same methods largely ineffective and frequently negative on live models. | Skip the claim |
| Keyword density and stuffing | Measured at −8% even on the synthetic engine that produced the friendliest numbers in this table. Nothing published since has rescued it. | Skip |
Any single measurement of this is noise.
This is the most commercially important fact in the category, and it is the reason I will not sell you a monthly visibility dashboard.
Identical prompts returned the same list of recommended businesses fewer than one time in a hundred, and the same list in the same order roughly one time in a thousand, across 3,000 runs. Google AI Overviews changed on average every 2.15 days across 43,000 keywords, with about 45% of cited sources new between consecutive observations. The same ChatGPT prompt run in minimal versus high reasoning mode overlapped on only 25.6% of cited domains.
So a single number from a single run tells you nothing, and a dashboard that moves month to month is mostly measuring its own sampling. What survives is a percentage taken across repeated runs and reported as a range, with the prompt set and the dates published alongside it; a technical readiness check, which is binary and stays true whatever the engines do next; and whether any of it turned into pipeline in your CRM.
What this rules out. A single monthly share-of-voice number reported as performance. Any promise of a “position” in an AI answer. A guaranteed placement, in any form. If someone offers you one of those, the number they are selling cannot be checked.
It will be wrong eventually. That is the category, not the page.
The share of AI Overview citations that also ranked in Google’s top ten halved in eight months. Anyone selling AEO off a fixed playbook is selling last quarter. So this ledger gets re-graded on the first Monday of each quarter: every row checked for a new controlled study, a replication or a retraction, and the changelog below updated whether or not anything moved.
If you find a study that changes one of these grades, send it to help@thepipelinefixer.com and I will re-grade the row and say who sent it.
// 8 September 2026 · First published. Nine tactics graded.
// Next re-grade due: first Monday of December 2026.
Every figure above, where it came from.
Seer Interactive, 2026: content recency and AI visibility (7,683 cited pages).
Ahrefs, 2025: do AI assistants prefer fresh content (17M URLs).
Kevin Indig, 2026, via Search Engine Land: where in the page ChatGPT cites from (18,012 verified citations).
Google: AI optimisation guide and AI features and your website.
Ahrefs, 2025: what correlates with AI brand visibility (75,000 brands).
Ahrefs, 2026: schema markup and AI citations (1,885 pages against matched controls).
Ahrefs, 2026: the llms.txt server-log study (137,210 domains).
Aggarwal et al., 2023: GEO: Generative Engine Optimization, the source of the 40% and 34% figures.
Puerto et al., NeurIPS 2025: C-SEO Bench, the replication that did not reproduce them.
SparkToro / Gumshoe, 2026, via Search Engine Land: AI recommendation lists rarely repeat (3,000 runs).
Ahrefs, 2025: AI Overviews change every two days (43,000 keywords).
Semrush, 2026: reasoning modes and citation overlap (100 prompts, two modes).
Where an original was paywalled, the named secondary report is linked instead. No provider, agency or tool vendor is named anywhere on this page: the tactics are graded, not the people selling them.
The measurement you can check, and the line in your CRM.
Four of the nine tactics above are worth doing and none of them take long. The hard part is knowing where you stand before and after, in a way that survives someone re-running it — and then seeing whether any of it reached a deal.
That is the Answer Audit, $1,250: a documented prompt set run three times across three engines, reported as a range with the raw log and the dates attached, plus the technical readiness half that stays true whatever the engines do next. I add it to Find the leak rather than sell it on its own, because the pipeline question usually matters more than the visibility one.
The four tactics marked Do are yours to act on for free. If you would rather I did them, each is its own line on the itemised list, bought after the Audit: crawler access $500, answer-first pages $950, one consistent entity $700, answer pages $1,100. Taken together, every one after the first comes off $130 for the shared setup.