What actually works in AI search? An AI SEO checklist sorted by evidence

Open LinkedIn in any given week and you'll find a new must-do for AI search. Add an llms.txt file. Put schema markup on every page. Rewrite everything as FAQs. Some of this advice is sound. Some of it is a guess presented as a rule. Very little of it tells you which is which.

‍

This GEO checklist sorts the most common AI SEO tips by how much evidence stands behind them. Each tactic is labelled best practice, early bet or one to watch, and grouped by what it helps AI do: find you, understand you, cite you or recommend you. Use it to build an AI search strategy around what's proven, test what's promising, and stop spending time on what isn't.

Why most AI search advice is hard to trust

The field is young, and the labels keep changing

Whether you call it AI SEO, LLM SEO, generative engine optimisation, answer engine optimisation or ChatGPT SEO, it's the same job. The term generative engine optimisation only entered academic research in late 2023, so most of the advice you'll read is less than three years old.

AI answers are hard to measure

When SparkToro ran the same 12 questions 2,961 times across three AI tools, there was less than a 1 in 100 chance of getting the same list of brands twice in 100 tries (SparkToro, 2026). A tip that 'worked' for someone after checking a few answers may just be normal variation.

Correlation is easy, cause is hard

Plenty of studies show that sites doing a certain thing get cited more. But those sites usually do many other things well too. Far fewer studies test what happens when a site changes one thing and nothing else.

Many tips come from people selling the tactic

That includes search tool vendors, and it includes us. Where a finding comes from a vendor study, we say so, and we treat it as directional rather than settled. A good AI search strategy starts by knowing the difference.

Best practice, early bet or one to watch: how we sorted the tactics

We use three labels, taken from Wordflow's AI Visibility Playbook:

‍

• Best practice. Backed by consistent evidence, or by how AI systems are documented to work. Do these first.
• Early bet. Sound logic and some supporting evidence, but not yet proven. Worth testing where it fits your priority.
• One to watch. Popular, but with no reliable evidence yet that it improves AI citations. Only do it if it's cheap or useful for other reasons.

‍

Each tactic also has an effort rating (low, medium or high) for your team's time. A low-effort best practice is your easiest win. A high-effort tactic marked one to watch is the first thing to cut.

‍

The checklist follows the four stages AI goes through before it recommends a brand, which Wordflow calls Found, Understood, Cited and Recommended. We explain the framework in our guide to getting AI to recommend your brand, not just mention it.

Tactics that help AI find you

The strongest evidence here is technical. 'Server-rendered' means a page arrives with its text already in it, rather than loading that text afterwards with code. That matters because when Vercel studied close to a billion monthly AI crawler requests, it found none of the major crawlers ran JavaScript (Vercel, 2024). If your product details only appear after scripts run, AI may never see them.

‍

Crawler rules matter for a similar reason. Cloudflare, which handles traffic for around 20% of the web, now asks every new domain whether to allow AI crawlers and blocks those without permission by default (Cloudflare, 2025). Check your settings on purpose, rather than discovering later that a security default shut AI out.

‍

Covering follow-up questions is best practice because of how some AI tools search. Google says its AI Mode breaks a question 'into subtopics' and searches for each one (Google, 2025). A page that also answers the questions around the main one has more chances to be pulled in. Our guide to query fan-out explains how to map those questions.

Tactics that help AI understand you

‍

AI builds its picture of your brand from many sources at once. If your website, LinkedIn page and directory listings each describe you differently, AI has to guess, and it often guesses wrong. A single approved description, used everywhere, is the cheapest best practice on this list.

‍

Wikidata is a free, structured database of facts about organisations, people and things. Adding an entry is logical, because many knowledge systems draw on it, but there's no direct evidence yet that it lifts AI citations. That makes it an early bet. Wikipedia has strict notability rules, so only pursue it if your brand genuinely meets them.

Tactics that get you cited

‍

This is where third-party evidence does the heavy lifting. AirOps analysed 21,311 brand mentions in AI answers and found that 85% came from third-party websites, and nearly 90% of those came from listicles, comparisons or reviews (AirOps, 2025). It's a vendor study, but the direction is consistent with other research: the websites AI already trusts in your category matter more than anything you publish yourself. Our guide to tracking AI citation sources shows how to find your Citation Core.

‍

Original evidence is the other best practice. Google's own guidance on generative AI search warns against 'commodity content' based on common knowledge, and recommends a unique, first-hand perspective instead (Google Search Central, 2026). Data that only you have gives publications a reason to cite you, and gives AI something it can't get from anyone else.

Tactics that get you recommended

‍

The best evidence for this stage comes from academic research. In a study presented at KDD 2024, researchers from Princeton and IIT Delhi found that adding statistics, quotations and citations to credible sources raised a page's visibility in generative engine responses by up to 40%, while keyword stuffing performed poorly (Aggarwal et al., 2024).

‍

Detailed reviews work on the same principle. A review that says 'great product' gives AI nothing to repeat. A review that says 'cut our month-end reporting from three days to one' gives AI a specific reason to recommend you for that use case. Tracking hedging, the caveats AI attaches to your name, tells you whether that evidence is landing.

Popular AI SEO tactics with weak evidence (so far)

Some of the most talked-about AI SEO tips have the least evidence behind them. That doesn't mean they're harmful. It means you shouldn't expect them to move your AI visibility on their own.

llms.txt: no measurable link to AI citations

This is a text file, proposed in 2024, that gives AI tools a summary of your site. SE Ranking analysed about 300,000 domains and found that around 10% had one, with no measurable link between having the file and how often a site was cited by AI (Search Engine Journal, 2025). Google's guidance is blunt: you don't need AI text files to appear in Google Search, and 'Google Search itself doesn't use them' (Google Search Central, 2026). Verdict: one to watch. Adding one is fine if it's quick, but don't expect results.

Schema markup: good housekeeping, not an AI citation tactic

Schema is code that labels the information on a page, such as a product's price or a company's address. Ahrefs tracked 1,885 pages that added schema and compared them with around 4,000 similar pages that didn't. AI citations barely moved: no meaningful change in two of the AI tools tested, and a small decline in Google's AI Overviews (Ahrefs, 2026). Google also says 'structured data isn't required for generative AI search'. Two caveats are worth knowing. The study only covered pages AI was already citing, and schema still helps with rich results in regular search. Verdict: keep it as good housekeeping, not as an AI search tactic.

Splitting content into tiny pieces: not required

Some advice says AI prefers short, fragmented pages. Google says there's 'no requirement to break your content into tiny pieces' and no ideal page length. Clear sections that make sense on their own are best practice. Fragmenting a good page into dozens of thin ones isn't.

‍

Notice the pattern. The tactics with weak evidence are quick, technical and easy to sell. The ones with strong evidence are harder: a readable website, a consistent story, independent coverage and detailed reviews.

How to choose tactics for your priority layer

Even the best practices on this list aren't all worth doing right now. The tactics that matter are the ones that fix the stage holding your brand back, which we call your priority layer. Here's how to choose:

‍

1. Find your priority layer. It's the first of the four stages that's clearly broken. Our guide Does AI Know Your Business Exists? shows how to check each one in order.
2. Start with that layer's best practices, lowest effort first. These are your most reliable wins.
3. Add one early bet only if it fits the same layer and you can measure whether it worked.
4. Leave tactics marked one to watch alone, unless they're cheap or useful for other reasons.
5. Limit yourself to two or three tactics per 90 days, each with a named owner and a deadline. Our guide to planning AI search engine optimisation across teams explains how to assign them.

‍

For example, if your priority layer is Cited, you might start by finding your Citation Core (best practice, medium effort) and grouping those sources by type (best practice, low effort), then move on to original research (best practice, high effort). llms.txt and schema wouldn't make the list, because they don't address the problem.

‍

This is the difference between a checklist and an AI search strategy. A checklist tells you everything you could do. A strategy tells you the few things worth doing now, in the right order.

How Wordflow turns your priority layer into a task list

Wordflow tracks the questions your buyers ask across up to 12 AI engines, depending on your plan, every week. It shows how your brand performs at each of the four stages, so you can see your priority layer instead of guessing it.

‍

Each part of the checklist maps to something you can check in Wordflow:

‍

• Found. The AI readiness check tests whether AI can read your site and gives you technical recommendations to work through.
• Understood. Brand Narrative campaigns compare what AI says about you with the approved description in your Brand Profile, so you can see where answers drift.
• Cited. Pages and Citations shows which websites AI relies on for your category, which is your Citation Core. Missed Prompts shows the questions where competitors appear and you don't.
• Recommended. Brand Comparison campaigns track how often AI recommends you and how often it hedges.

‍

The GEO Action Hub then turns those findings into a prioritised task list for your priority layer, giving your AI search strategy a clear starting point, so you work on what your own data shows, not on the latest tip from social media. When content is the right fix, GEO Writer helps you create it.

‍

Ready to find out which tactics your brand actually needs? Start with Wordflow, or book a demo and we'll walk you through it. To learn the full approach to AI SEO at your own pace, download the AI Visibility Playbook.

‍