Evidence review
What actually works, graded by evidence
Sorted by how well supported each tactic is rather than how often it is repeated. Three tiers: demonstrated in published research, consistent with the data but unproven, and asserted with no evidence we can find. A good deal of what this industry sells sits in the third tier.
Tier 1
Demonstrated in published research
One peer-reviewed study underpins most defensible claims in this field. It tested content modifications across roughly 10,000 queries and measured the change in visibility inside generated answers.
Targeted content changes raised visibility in generative engine responses by up to 40%.
GEO-bench: roughly 10,000 queries across nine datasets. Authority-style edits, adding citations, statistics and quotations, outperformed cosmetic edits such as rewriting, simplification and keyword work. Effect size varied by domain.
The consistent pattern: authority signals outperform cosmetic edits. Adding citations, statistics and quotations moved visibility materially. Rewriting for fluency, simplifying, and keyword optimisation did much less. Keyword stuffing: the reflex of two decades of SEO, performed worse than leaving the page alone.
Replace vague claims with specific figures
'Delivered in 34 markets since 2019' rather than 'extensive international experience'. The best-evidenced single change available, and it works on pages you already have.
Attribute claims to named, dated sources
A model can safely quote a sourced statement. An unsourced assertion is a liability it has no reason to repeat.
Include quotable expert statements
Attributable quotations from identifiable people were among the effective interventions.
Keyword density optimisation
Measured as actively counterproductive for citation. This is the clearest case where classical SEO instinct works against you.
Read the paper rather than the summaries. Per-technique percentages circulate widely, differ between write-ups, and do not all appear in the abstract. The safe claim is directional: evidence density beats stylistic polish.
Tier 2
Consistent with the 2026 data, not formally proven
These follow from measurement of how these systems behave. They are reasoned from observation rather than tested experimentally, and they are labelled as such.
Answer-first structure
Selection happens at passage level, and over half of sampled AI Overview citations came from the first 30% of the cited page. Stating the question as a heading and answering it in the opening lines gives the system something liftable.
55% of sampled AI Overview citations came from the first 30% of the cited page.
Freshness maintenance
For commercial and evaluation queries, 83% of citations came from pages updated within twelve months. Systematically revising existing high-intent pages is cheaper than publishing new ones and better supported.
For commercial and evaluation-stage queries, 83% of AI citations came from pages updated within the previous 12 months, and over 60% from pages refreshed within six months.
Third-party presence over own-domain publishing
Roughly 85% of AI references point at third-party platforms rather than brand-owned sites, and the top fifteen domains absorb about 68% of all citations. For most organisations the achievable goal is accurate representation inside sources the engines already trust, not becoming one of them.
In practice: community threads where your category is discussed, reference entries, trade coverage, and comparison content on established publications. Uncomfortable for teams whose entire budget points at their own domain.
Roughly 85% of AI references point to third-party platforms rather than brand-owned sites.
Entity resolution before content volume
If a system cannot determine that your brand name, legal name and domain are one organisation, published content cannot be attributed to a business. Consistent naming, alternateName in structured data, a reference-source entry and corroborating third-party profiles are the mechanism.
Language coverage
An assistant answering in Japanese draws on Japanese-language sources. English-only content is structurally invisible on that surface. Inferred from how retrieval works rather than directly measured.
Tier 3
Sold confidently, with no evidence behind it
Not claims that these are harmful, claims that nobody has shown they work.
llms.txt
The clearest case, because it has now been measured directly.
A study of 300,000 domains found llms.txt on 10.13% of them; a separate study of 137,000 sites found 97% of llms.txt files received zero traffic in May 2026.
Google documentation updated in June 2026 states llms.txt has no effect on Search rankings or AI Overviews. John Mueller noted that no AI crawler has claimed it extracts information from the file.
Google's documentation states it has no effect on Search or AI Overviews. John Mueller's argument against it is the more interesting one: a self-reported manifest cannot work as a quality differentiator, because every site would claim to be the best one. Publish it if you like. It is cheap and harmless, but it is housekeeping, not strategy.
Everything else in this tier
Guaranteed placement in AI answers
Nobody controls organic citation inside a generated answer. A supplier promising a position is selling something they do not control.
Schema markup as a direct citation lever
Structured data helps parsing and matters for entity resolution, both prerequisites. The specific claim that adding FAQPage markup raises citation rates is, as far as we can find, unmeasured.
A single 'AI visibility score'
Retrieval and model memory move on different timescales for different reasons. One number averaging them cannot be acted on.
Prompt-injection tricks in page text
Hidden instructions aimed at models. Ineffective at scale, trivially detectable, and a reputational liability when found.
Publishing volume for its own sake
Volume without extractable structure and verifiable specifics produces pages nothing quotes, and the research shows undifferentiated keyword-heavy content scoring negatively.
Sequence
A defensible order of work
Ordered by dependency, not by importance. Several of these fail if attempted out of sequence.
-
01
Baseline before publishing
Ask the questions your buyers ask, in the assistants they use, with search on and off. Record who gets named. Without this there is no way to distinguish progress from model drift.
-
02
Fix entity resolution
Make it unambiguous that your brand name, legal name and domain are one company, and that something on the open web states your category. Produces no visible output; skipping it makes everything downstream unattributable.
-
03
Rewrite existing high-intent pages
Question as heading, answer in the opening lines, specific figures, named sources. Cheaper than new content and better evidenced.
-
04
Add verifiable specifics everywhere
The strongest single intervention in the published research. Costs editorial discipline, nothing else.
-
05
Work on third-party representation
Given that most citations point elsewhere, accuracy in the sources engines already trust matters more than your own publishing cadence.
-
06
Re-measure, separating the two surfaces
Expect retrieval to move in weeks and memory not to move for a long time. Reporting them together is what makes working programmes look broken.
The uncomfortable part: steps one and two produce nothing publishable and no visible activity, and they are the ones most often skipped in favour of shipping content that then cannot be attributed to anyone.
Measurement
How to measure without fooling yourself
- Fix the question set. Build it from the buying journey, not from search volume. Changing the questions between runs destroys comparability.
- Run both modes. Search enabled and disabled, every time. They answer differently and mean different things.
- Track mention separately from citation. Being described without a link still shapes a shortlist and is invisible to link-counting.
- Keep control questions. Track a category you do not sell into, so you can tell a real change from the models shifting underneath you.
- Pin model versions. Upgrading the model silently invalidates every earlier reading.
- Expect noise in position. Presence or absence is robust; exact rank is not.
- Do not expect traffic. AI Overview citations are clicked around 1% of the time. Citation is an influence metric, not an acquisition channel.
When an AI Overview is shown, users click one of its cited source links about 1% of the time.
Sources
Where these numbers come from
Every figure on this page is external and linked. The studies measure different populations over different windows, so read them as converging evidence rather than one authoritative number.
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande, ACM SIGKDD 2024Targeted content changes raised visibility in generative engine responses by up to 40%. arxiv.org ↗
- Omnibound, AEO statistics compilation 202655% of sampled AI Overview citations came from the first 30% of the cited page. www.omnibound.ai ↗
- Omnibound, AEO statistics compilation 2026For commercial and evaluation-stage queries, 83% of AI citations came from pages updated within the previous 12 months, and over 60% from pages refreshed within six months. www.omnibound.ai ↗
- Omnibound, AEO statistics compilation 2026Roughly 85% of AI references point to third-party platforms rather than brand-owned sites. www.omnibound.ai ↗
- AI Platform Citation Source Index 2026The top 15 domains account for roughly 68% of all citations produced by the five major AI answer engines. everything-pr.com ↗
- SE Ranking and Ahrefs studies, reported by Digital AppliedA study of 300,000 domains found llms.txt on 10.13% of them; a separate study of 137,000 sites found 97% of llms.txt files received zero traffic in May 2026. www.digitalapplied.com ↗
- Google Search RelationsGoogle documentation updated in June 2026 states llms.txt has no effect on Search rankings or AI Overviews. John Mueller noted that no AI crawler has claimed it extracts information from the file. baselinelabs.ai ↗
- Pew Research CenterWhen an AI Overview is shown, users click one of its cited source links about 1% of the time. searchengineland.com ↗
Keep reading
FAQ
Common questions
What is the single highest-value change?
Replacing vague claims with specific, sourced facts. Strongest published evidence, works on pages you already have, costs only editorial discipline.
Does llms.txt help?
There is no evidence it does. Google states it has no effect on Search or AI Overviews, no major AI provider has committed to reading it, and a study of 137,000 sites found 97% of llms.txt files received zero traffic in May 2026. Cheap and harmless; not a lever.
Should I publish more content?
Usually not first. Volume without extractable structure and verifiable specifics produces pages nothing quotes, and undifferentiated keyword-heavy content measured negatively. Revising existing high-intent pages is better evidenced and cheaper.
How long before this shows up in measurement?
Retrieval-surface changes can appear within days to weeks. Model-memory changes depend on retraining and third-party coverage, so months to years. One timeline quoted for both is the most common way this gets mis-sold.
Can I pay to appear in AI answers?
Not in the organic citation. Advertising placements exist alongside AI answers on some surfaces, but citations are earned through content, structure and third-party corroboration.
Does schema markup improve citations?
It helps machines parse your page and matters for entity resolution, both of which are prerequisites. The narrower claim that it directly increases citation rate is not well evidenced. Treat it as hygiene.
Want help getting into AI answers? <span class='tone'>Talk to Lifewood Data Technology.</span>
Lifewood Data Technology runs AEO and GEO programmes for enterprise brands. We measure where you stand first, then fix what is blocking you.
- 100+ languages
- 50+ delivery centres
- Measured monthly
Lifewood Data Technology publishes this reference.