What actually works in AEO in 2026, and what doesn't

How to read this ledger

Every entry has four parts: the evidence, the source, how strong that evidence is, and what we would do about it. We rate the evidence on three levels.

• Official The platform's own documentation, data or announcement. We treat it as fact and re-check it before acting on it.
• Large study A big dataset, survey or controlled test. Most are correlational: they show that strong brands do something, not that doing it makes a brand strong. We treat them as directional.
• Vendor claim One company measuring a market it sells into. Often useful, rarely replicated. We name the vendor and never use it as a headline on its own.

Many large studies are also published by tool vendors. Where that's the case, we say so.

The AEO evidence ledger at a glance

What works in AEO in 2026?

Seven levers have evidence behind them. They share one idea. AI assistants repeat what the sources they trust say about you, and they quote the pages that answer a question most cleanly.

1. Earned media: roundups, trade press and community mentions

Verdict: works. It is the best-evidenced lever in AEO, and it is the centre of our programme and retainer.

• Evidence: In Muck Rack's analysis of 25 million links cited by AI, 84% were earned media and 27% were journalism. Press releases made up 1.1% and paid placements 0.3%. In Ahrefs' study of 75,000 brands, unlinked brand mentions correlated with AI visibility at about 0.66 to 0.71, against 0.22 for backlinks. Reddit, YouTube, LinkedIn and Wikipedia lead the domains AI assistants cite (Peec AI, 30 million sources).
• Source: Muck Rack, May 2026. Ahrefs, 26 May 2025 and 12 December 2025. Peec AI, early 2026.
• Strength: Large study All three samples are very large. All three come from companies that sell PR, SEO or tracking software, and the Ahrefs figures are correlations.
• What to do: Find the roundups and comparison articles that AI assistants already cite in your category, and get onto at least three of them. Turn awards, data and customer results into stories a trade editor will run. In B2B, complete your LinkedIn company page. Take part in Reddit threads openly, as a named employee, but treat it as a Google AI Mode and Perplexity tactic: it stopped working for ChatGPT in August 2026.

2. YouTube, long form

Verdict: works. It is the strongest single signal measured so far.

• Evidence: In the same Ahrefs study, YouTube mentions had the strongest correlation with AI visibility of any factor measured: about 0.74 across ChatGPT, AI Mode and AI Overviews. Otterly's study of AI-cited videos found 94% were long form. Views, likes and subscribers made no difference, and longer descriptions helped.
• Source: Ahrefs, 12 December 2025. Otterly, March 2026.
• Strength: Large study Correlation, not proof. Brands that are talked about on YouTube tend to be strong brands anyway.
• What to do: Publish two long-form explainers that answer the same questions as your answer pages, then one a quarter. Put your brand and the category in the title. Add chapters and a long, specific description. Don't chase views: they don't predict citations. This is a briefing job, not a film production.

3. Review platforms, chosen by sector

Verdict: works, and it matters most in B2B software and local services.

• Evidence: In SE Ranking's study of 30,000 commercial software searches, a third of AI Overviews cited a review platform. Five platforms took 88% of those citations: Gartner Peer Insights 26.0%, G2 23.1%, Capterra 17.8%, Software Advice 12.8% and TrustRadius 8.3%. In G2's 2026 buyer behaviour report, review sites (38%) and AI chatbots (37%) now shape B2B software shortlists in equal measure.
• Source: SE Ranking, 29 January 2026. G2, 22 July 2026, a survey of B2B software buyers.
• Strength: Large study A large citation study and a large survey. The survey comes from a review platform that benefits from the finding.
• What to do: Pick the platforms your sector is judged on. For software, claim G2 and Gartner Peer Insights first. For local businesses, Yelp, Tripadvisor and Facebook. For retail, Trustpilot and your product ratings. Check you're filed in the right category, and ask for a steady flow of recent reviews: recency matters more than the total.

4. Answer-first content that is genuinely maintained

Verdict: works. This is the part of AEO that lives on your own site.

• Evidence: Seer Interactive found that 75% of 7,683 pages cited by AI had been updated within the last year. Over a quarter of them were first published two or more years earlier: they earned their freshness by being maintained. A 2026 survey of 45 studies found topical relevance is the most reproducible lever. It also found the "up to 40%" lift from adding statistics and quotes, from the original GEO paper, holds only in a lab setting.
• Source: Seer Interactive, 24 July 2026. Martinez survey, arXiv, 15 July 2026. GEO paper, KDD 2024.
• Strength: Large study Vendor claim The survey is academic. Seer is a search agency. The statistics-and-quotes gain is proven in the lab and unproven in live AI assistants.
• What to do: Use question-style headings that match what buyers ask. Put the direct answer in the first two sentences, and add a visible question-and-answer block. Cite your sources, but never bury the answer under them. Update the pages AI assistants already cite with real changes, not a new date. Cover the sub-questions too, because assistants split each question into several background searches.

5. Comparison, "best" and pricing pages

Verdict: works. It is the format AI assistants cite most when a buyer is choosing.

• Evidence: In Overthink Group's analysis of B2B software citations, 70.8% linked to pages with "best", "top" or a similar word in the title, and 51.6% included "2026". Google AI Overviews and Gemini cited vendors' own pages heavily; ChatGPT and Perplexity leaned on third parties. In Semrush's study of 1,094 categories, only 15.2% had a clear owner in ChatGPT, and owning one meant appearing for four of five question types.
• Source: Overthink Group, 14 July 2026. Semrush, 20 July 2026.
• Strength: Vendor claim Mid-sized samples from an agency and a software vendor, one of them gathered over a single week.
• What to do: Publish honest comparison pages with the year in the title and real prices on the page. Add one "alternatives to" page for each category leader, and a public pricing page if you can. Cover all five question types for your category: what it is, how options compare, the alternatives, the use cases and the buying question.

6. Answer crawler access and readable HTML

Verdict: works. It is the entry ticket, not the whole game.

• Evidence: OpenAI says sites that block OAI-SearchBot will not be shown in ChatGPT search answers. From 15 September 2026, Cloudflare blocks its Training and Agent crawler categories by default on ad-bearing pages for new zones, and it classes ChatGPT-User as an Agent. Google says there are no additional requirements for AI Overviews or AI Mode beyond being indexed and eligible for a snippet. Vercel and MERJ found that the main AI crawlers other than Googlebot do not run JavaScript.
• Source: OpenAI crawler documentation. Cloudflare, 1 July 2026, effective 15 September 2026. Google Search Central. Vercel and MERJ, December 2024.
• Strength: Official Large study The crawler rules are the platforms' own documentation. The JavaScript finding is a large log study, and no newer one has replaced it.
• What to do: Allow the answer crawlers in robots.txt, then check that your CDN or firewall lets them through as well. Serve your words in the HTML, not only after JavaScript runs. Keep the sitemap complete, consolidate duplicate pages, and make sure nothing important is set to noindex or nosnippet. If you have an SEO agency, this is their ground, and we work alongside them.

7. Google Business Profile for local, merchant feeds for retail

Verdict: works, where your sector is read from a profile or a feed.

• Evidence: In BrightLocal's study of 1.97 million local AI citations, Google Business Profile took 28.6%, second only to the businesses' own websites. Google tells businesses to keep their Business Profile and Merchant Center current to appear in its AI features. For retail, the feed is now the main route into AI shopping answers. Google Merchant Center free listings reach AI Mode and Gemini, and Microsoft Merchant Center reaches Copilot. OpenAI's product feed reaches ChatGPT, but currently targets the US.
• Source: BrightLocal, 16 September 2026. Google Search Central. OpenAI commerce feed specification.
• Strength: Official Large study The feed routes are vendor documentation. The local figure is a large citation study.
• What to do: For local businesses, complete one Business Profile per branch, with the right categories, services, photos and recent reviews. For retailers, keep one clean feed with GTINs and submit it to Google and Microsoft Merchant Center, then to OpenAI where your market is eligible. Apple Business Connect and Bing Places are cheap to claim, but unproven.

What doesn't work in AEO?

Six things are sold hard and supported weakly. Google's own guidance rules out three of them directly. One of them is different: it is hygiene, worth keeping because it is cheap, but never a reason to buy a programme.

1. llms.txt

Verdict: doesn't work. No major AI provider has been shown to use it.

• Evidence: Ahrefs' log data found that 97% of 137,210 llms.txt files got no requests at all in May 2026. SE Ranking's October 2026 study of 300,000 domains found 10% of sites have one, and removing it from their prediction model made the model more accurate. Google says you don't need "new machine readable files, AI text files, markup, or Markdown" to appear in its AI features.
• Source: Ahrefs, reported by Search Engine Journal, June 2026. SE Ranking, 15 October 2026. Google Search Central, 10 July 2026.
• Strength: Official Large study
• What to do: Nothing. If you already have one, check it isn't broken or full of unedited AI output, because anyone can read it. Don't pay anyone to write one.

2. Schema added "for AI", FAQPage markup included

Verdict: doesn't work as an AI citation lever.

• Evidence: Ahrefs tracked 1,885 pages that added JSON-LD structured data and compared them with about 4,000 that didn't. The change in AI citations on ChatGPT and AI Mode was statistically indistinguishable from zero. Google says there is no special schema.org markup you need to add. No independent study shows FAQPage markup changes citations: the visible questions and answers are what retrieval reads.
• Source: Ahrefs, May 2026. Google Search Central.
• Strength: Official Large study
• What to do: Keep an accurate Organization block and the schema you have, consistent with the visible page. It is hygiene, and it still supports classic search features, which is why this page has it too. Put your effort into the visible Q&A text, not the markup. Don't buy an "AI schema" package.

3. Hygiene sold as a lever: sameAs, Wikidata, IndexNow, title and description

Verdict: hygiene. Keep it, because it is cheap and harmless, but don't buy it as a reason to be cited.

• Evidence: We found no study that links sameAs links, a Wikidata item, or a title tag and meta description to AI citations. The schema test above covers sameAs. IndexNow reaches Bing, Amazon, Naver, Seznam, Yandex and Yep: Google is not a participant, and no AI assistant is on the list. Wikipedia is a top source for ChatGPT and Gemini, but most companies don't meet its notability bar, and Wikidata alone has no measured effect.
• Source: IndexNow FAQ. Ahrefs, May 2026.
• Strength: Official The IndexNow participant list is the protocol's own. For the rest, the evidence is an absence of evidence.
• What to do: Spend an afternoon on them once. Write a title and description that name your category, and add sameAs links to your real profiles. Correct your Wikidata item if you have one. Switch on IndexNow if you publish often or rely on Bing and Copilot. Then stop. Anyone who quotes these as the reason AI assistants will cite you is selling hygiene as a lever.

4. Chunking or rewriting copy "for AI"

Verdict: doesn't work, and it can hurt.

• Evidence: Google's guide says: "There's no requirement to break your content into tiny pieces for AI to better understand it." Its systems understand pages that cover several topics. A 2026 survey of 45 studies found that citation-oriented rewrites can impair retrieval, and that no technique has a stable, cross-platform effect over time.
• Source: Google Search Central, 10 July 2026. Martinez survey, arXiv, 15 July 2026.
• Strength: Official Large study
• What to do: Write for the buyer. Answering a question directly in the first two sentences is good writing, and it helps people and machines alike. Slicing pages into fragments or rewriting them in a "robot" register is a different thing, and you don't need it.

5. Wire releases and paid placements

Verdict: doesn't work, and paid mentions carry risk.

• Evidence: AI answers cite coverage, not distribution. In Muck Rack's analysis, press releases made up 1.1% of cited links and paid or advertorial content 0.3%. Google's guide states: "Seeking inauthentic 'mentions' across the web isn't as helpful as it might seem."
• Source: Muck Rack, May 2026. Google Search Central, 10 July 2026.
• Strength: Official Large study
• What to do: Spend the wire and advertorial budget on one story a trade editor wants to run: new data, a customer outcome, or an award with a real angle. For a company with certifications, regulated products or enterprise clients, a paid-mention scheme is a reputational risk with no measurable upside.

6. Chasing one platform's citation diet

Verdict: doesn't work, because the diet changes overnight.

• Evidence: In Promptwatch's data, Reddit's share of ChatGPT Search citations averaged 3.83% from 18 July to 7 August 2026. It then fell to 0.52% from 14 to 17 August: an 86% relative drop, caused by a change in how ChatGPT builds its background searches. A month later, Peec AI found ChatGPT search running on its own index rather than mainly on Bing's. Reddit still holds close to half of Perplexity's top-ten citation share (Profound). Growth Memo found that 91% of citations appear in only one AI assistant.
• Source: Promptwatch, 20 August 2026. Peec AI, 4 September 2026. Profound. Growth Memo.
• Strength: Large study Vendor claim The figures come from tracking vendors and an industry newsletter, and they agree on the direction.
• What to do: Spread your presence across media, YouTube, review platforms and your own pages. Treat Reddit as a Google AI Mode and Perplexity tactic, not a ChatGPT one. Measure each AI assistant separately every month, with several runs per question, so a platform change shows up as a platform change and not as your failure.

What changed in AEO between May and October 2026?

Seven dated changes between May and October 2026 shaped this edition of the ledger.

• 13 May 2026. GA4 gains a default "AI Assistant" channel. Google's list names ChatGPT, Gemini, DeepSeek, Copilot and Grok, but not Perplexity or Claude. It excludes AI Overviews and AI Mode, and it is not retroactive (Google Analytics Help).
• 10 July 2026. Google updates its AI optimisation guide: optimising for AI search is "still SEO", with no special schema, no AI text files and no chunking needed (Google).
• 8 to 17 August 2026. ChatGPT changes how it builds its background searches, and Reddit's share of its citations falls 86% (Promptwatch, vendor research). Community presence becomes a Google AI Mode and Perplexity lever, not a ChatGPT one.
• 31 August 2026. Search Console's Generative AI performance report reaches every site. It shows impressions in AI Overviews and AI Mode combined, by page, country, date and device, with no clicks or queries (Search Console Help, Search Engine Journal).
• 4 September 2026. ChatGPT search is found running on OpenAI's own index, with search-result scraping as a fallback and Bing used mainly in Deep Research. Being in Bing no longer means being in ChatGPT (Peec AI, vendor research).
• 15 September 2026. Cloudflare blocks its Training and Agent crawler categories by default on ad-bearing pages for new zones, while Search stays allowed. ChatGPT-User counts as an Agent (Cloudflare).
• 13 October 2026. Bing Webmaster Tools adds Topics Portfolio, in beta: the grounding queries where competitors are cited and you are not (Bing Webmaster blog). Like the rest of Bing's AI Performance report, it covers Copilot, Bing's AI summaries and unnamed partners, so it is not ChatGPT data.

How we use this ledger in client work

This ledger is the "what not to spend money on" section of every AI Visibility Audit we run. It's also why our programmes spend most of their time on roundups, trade press, reviews, YouTube, comparison pages and answer content, and very little on files and markup. Organization schema, sameAs, Wikidata, IndexNow and title tags stay on our client board as hygiene steps: checked automatically, and never sold as levers.

The foundations are SEO, and Google says so. We do the parts a normal SEO retainer usually doesn't: roundups, reviews, PR, video briefs, crawler access through the CDN, and measurement of each AI assistant separately. If you already have an SEO agency, we work alongside it rather than replace it.

When will this ledger be updated?

Every quarter. The next scheduled review is December 2026. When a platform changes its behaviour between reviews, we add a dated note here rather than wait. If you think we've missed something or got something wrong, email hello@helloworld.ie with the source.