The AI-readability checklist: 12 things to fix this week
A short, opinionated list of fixes that change how AI crawlers see your site. Most of them are boring, and every one counts.
TL;DR. A 12-item AI-readability checklist for B2B sites: allow the search crawlers you need in robots.txt, server-render high-intent pages, ship clean Organization / Product / FAQPage schema, build canonical competitor comparison pages, use semantic HTML for tables, tighten your entity description across the open web, write the answer in the first paragraph, earn third-party citations, and monitor AI visibility weekly. You can finish them in a sprint. None of them is a clever trick, and Google’s own guide tells you to skip the “AEO/GEO hacks.”
Most “AI optimization” advice is either too vague to be useful (be helpful, write good content) or too clever to be true (write Python prompts in your robots.txt to bribe ChatGPT). The real work is a list of small, dull fixes nobody gets credit for shipping. Google’s own guidance points the same way: there are no special requirements to appear in AI Overviews beyond normal SEO best practice (Google).
Here are the twelve fixes. Each one checks something a crawler or an AI feature actually reads, with the vendor doc behind it where one exists. Work top to bottom. Most teams can clear the list in a week if they make it a priority.
Allow the search crawlers in your robots.txt
The most expensive mistake here is a blanket disallow for “AI scrapers” that also blocks the crawlers used for search answers. Split the list in two. Search answers: OAI-SearchBot for ChatGPT search, and Googlebot for Google Search, including AI Overviews and AI Mode (OpenAI; Google). Training and grounding controls: GPTBot and Google-Extended. Blocking those doesn’t by itself remove you from ChatGPT search or Google Search (Google). Review ClaudeBot and anthropic-ai (Claude), PerplexityBot and Applebot-Extended against each vendor’s docs. Open your robots.txt now. If a search crawler you care about is disallowed, fix that first. OpenAI says a robots.txt change can take about 24 hours to reach ChatGPT search.
Server-render your high-intent pages
If your pricing, comparison, and product pages are rendered client-side and only assembled after a JS bundle runs, the AI crawler is probably seeing a near-empty page. Some bots execute JS now. Most still prefer rendered HTML, and here “prefer” means “downrank everything else.” Try the View Source test: load your highest-intent page, view source, and search for the most important sentence on the page. If the sentence is not in the HTML, fix that page. The other pages can wait.
Add a clean Organization schema with sameAs
Organization schema is what tells AI engines that the various names you go by are the same entity. The sameAs array should include your LinkedIn page, your Crunchbase profile, your Wikipedia entry if you have one, and any major directory listing. Two failures are common: no Organization schema at all, and Organization schema that contradicts the visible page. Google’s guidelines say not to mark up content readers can’t see (Google). The fix is one block of JSON-LD in your site-wide head. Validate it.
Put Product or SoftwareApplication schema on your product pages
If you sell software, Product / SoftwareApplication schema is where you put your name, description, category, pricing, and aggregate rating in machine-readable form. Google says not to mark up content readers can’t see (Google), and that structured data isn’t required for its AI features (Google). So use it to state your plans and prices unambiguously, and keep the same numbers in visible text where any crawler can read them.
Use real FAQPage schema, sparingly
FAQPage schema is one of the most overused tools in SEO, and since August 2023 Google shows FAQ rich results only for well-known, authoritative government and health sites (Google). Use FAQPage schema only on pages where you have real, buyer-asked questions, with answers a buyer would actually want. Three real FAQs beat fifteen synthetic ones.
Build one canonical comparison page per major competitor
“X vs Y” pages answer the exact comparison buyers ask AI engines, when they are well done. Well done means: a real table at the top, semantic HTML headers (use <th>, not <div> styled to look like one), explicit pricing, an honest assessment of where the competitor is better, and a clear summary at the bottom. The mediocre version of this page (a marketing-speak listicle that says you are best at everything) does not get cited. The honest version does.
Use semantic HTML on data-heavy pages
This sounds like 2008 advice, but it still holds. AI engines pull comparison tables, pricing tables, and feature matrices into their answers very directly. They pull cleanly when the table is a real <table> with <thead> and <th>, and badly when the table is a stack of divs styled into a grid. The fix is one engineering ticket, and every engine benefits.
Tighten your entity description across the open web
Look up how your brand is described in five places: your own about page, your LinkedIn company page, your Crunchbase entry, your G2 / Capterra profile if you have one, and the first sentence of your most-cited press piece. If those five descriptions disagree, the AI engine is reading a fragmented entity, and the fragmentation is costing you citations. Pick one canonical sentence (10 to 25 words) and use it everywhere. It’s dull work, and it’s worth a sprint.
Build citation hooks: definitions, named frameworks, statistics
A content pattern that makes a passage easy to attribute is “this thing has a name and the name is on this page.” Definitions of category-specific terms, frameworks you have named (a “[Brand] [Method]” or similar), original statistics from your own data: each one gets cited because each one is easy to attribute. If you have a number, give the number a sentence of its own. If you have a framework, give the framework a name and a page. Each one gives the model something to grab. (More on what makes a passage attributable in How AI engines decide which brands to cite.)
Write the answer in the first paragraph
Old SEO content was structured to keep readers on the page (delay the answer, build anticipation). AI synthesis works differently. The model is going to chunk your page, embed those chunks, and retrieve the chunk that most directly answers the prompt. If your answer sits in paragraph nine, that one paragraph has to carry the page. If your answer is in paragraph one, every chunk in the page benefits from the context. Lead with the answer. Add the nuance after.
Get three to five real third-party reference mentions per quarter
AI engines increasingly prefer sources that are not from the brand itself. If your own domain is the only place you get cited, you’re leaving AI visibility on the floor. “Do PR” won’t fix it. Getting into the structured listicles, comparison articles, and category roundups that buyers (and AI engines) reference will. One well-placed listicle on a high-authority third-party site does more than ten guest posts.
Set up monitoring so you know when something breaks
Nobody warns you that AI visibility can drop overnight. A schema change goes live and stops validating. A bot gets accidentally re-blocked in a robots.txt update. A retrained model decides it likes your competitor more this week. If you aren’t measuring, you find out three months later, after pipeline has already softened. Track AI visibility, share of voice, citation share, and sentiment on a weekly cadence at minimum. (For the case that AI visibility is the number to track above all, see AI visibility is the new market share.)
Should you still write FAQ-style content for AI?
The conventional advice you will see on most “GEO checklist” articles is “write FAQ-style content for AI.” The advice made sense in 2024. In 2026, FAQ-block farming is one of the lowest-return things on the list, and overdoing it can actively hurt you. Google’s guidance for its AI features asks for unique, non-commodity content and says no special markup is needed (Google). The engines are getting better at telling content written for buyers from content written to game them, and anything that reads as written for the bot will age badly. Write for buyers, mark up the structure cleanly, and the bots will follow. (More on the shift from rank-thinking to answer-thinking in The end of keyword rankings.)
What is missing from this 12-item list?
This is a 12-item list because the goal is something a team can actually finish. The full Zaraftis site audit goes further, mostly with variations on these themes (deeper structured data validation, per-page accessibility checks, more granular schema choices, citation graph mapping, prompt-level coverage testing) and a few that are highly specific to vertical or stack. Do these twelve well and what’s left is ongoing work rather than another sprint.
There’s no clever trick, just a list you can finish in a sprint. Brands that work through it look prescient two quarters later. Brands that wait for a fancier playbook will get one, it will cost more, and it will start with these twelve items anyway.
Frequently asked questions about AI-readability
Q: Which AI bot is most important to allow in robots.txt?
A: There is no single most important bot, because each company uses separate crawlers for search answers and for training. For ChatGPT search, allow OAI-SearchBot: OpenAI says sites that block it won’t be shown in ChatGPT search answers. GPTBot is OpenAI’s training crawler, and blocking it doesn’t remove you from ChatGPT search. For Google, normal Googlebot rules govern Search, including AI Overviews and AI Mode; Google-Extended is a training and grounding control for Gemini Apps and Vertex AI and doesn’t affect inclusion in Google Search. Also review ClaudeBot and anthropic-ai (Claude), PerplexityBot and Applebot-Extended against each vendor’s current docs. Match each disallow to the product you mean.
Q: Do AI engines actually execute JavaScript?
A: Some do, some do not, and the ones that do still prefer server-rendered HTML. Treat client-only rendering on a pricing or product page as a citation-loss bug. The simplest check is to open View Source on your highest-intent page and search for the page’s most important sentence. If it is not there, the AI crawler is not seeing it.
Q: Should every page have FAQPage schema?
A: No. Google says structured data isn’t required for its AI features, and since 2023 it shows FAQ rich results only for a small set of well-known government and health sites. Use FAQPage only on pages with real buyer-asked questions and real answers. Three real FAQs beat fifteen synthetic ones.
Q: How do I know if my entity description is fragmented?
A: Look up how your brand is described in five places: your about page, your LinkedIn company page, your Crunchbase entry, your G2 / Capterra profile, and the first sentence of your most-cited press piece. If those five descriptions disagree, the AI engine is reading a fragmented entity. Pick one canonical 10-to-25-word sentence and use it everywhere.
Q: How often should I re-check this list?
A: Run the technical checks (robots.txt, server-render, schema validation) any time you ship a site-wide change to those layers. Re-audit the off-site entity layer (LinkedIn, Crunchbase, G2) once a quarter. Track AI visibility, share of voice, citation share, and sentiment on at least a weekly cadence. A regression can show up overnight.
Q: How long until these fixes show up in AI visibility numbers?
A: No engine publishes a timetable. OpenAI says a robots.txt change can take about 24 hours to reach ChatGPT search. For everything else, re-scan at day 14 and day 28 and compare against your own baseline.
Want the full audit?
Zaraftis runs a site audit and turns the gaps into a ranked Actions board, not a 90-page PDF. The 7-day free trial needs no card: one brand, 15 prompts and one scan on ChatGPT. Prefer to talk it through? Book a call and we’ll show you what the engines say about you.