Which pages actually get you cited by AI engines?

Across five weeks of daily tracking, the pages AI engines actually pulled from our site were listicles, comparisons, and the homepage. Nine of our ten most-retrieved pages were one of those three. Not one long-form essay reached the top ten, and we've published more than 250 of them. If you're deciding what to publish next to get recommended inside a chat window, that ratio is the finding, and it cost us four months of writing to learn.

Here's the page-level ledger from our own instrument: what the engines retrieved, what they quoted, what they read and ignored, and the one thing the quoted pages share. We've written the full playbook of how we got recommended elsewhere. This is the autopsy underneath it.

Which of our pages do AI engines actually pull?

We track 50 buyer prompts daily across ChatGPT, Gemini, Perplexity, Claude, and Google AI Overview, measured against a 13-brand competitive field. Between July 13 and August 16, 2026, this is what the engines pulled from pitchkitchen.com most often.

The healthtech list is the standout. It gets a third of the traffic of our top page and earns 2.84 citations per conversation, the best rate of anything we publish at volume. Narrow beats broad. A page that says which firms fit a digital health company selling to clinical and economic buyers gives an engine something specific to repeat, and specificity is what gets repeated.

Why do some retrieved pages never get quoted?

Retrieval and citation are separate events, and the gap between them is where most content dies. An engine pulls your page into context, reads it, finds nothing worth lifting, and answers from someone else's page. You never appear. Our own tail looks like this:

• An essay of ours was pulled into 33 conversations and quoted zero times.
• Our press and podcast page was retrieved 40 times and cited once, a rate of 0.03.
• Our frameworks index was retrieved 37 times and cited twice.
• A well-written post on timing a positioning fix was retrieved 34 times and cited 4 times.

Compare those to a plain pricing explainer that only got retrieved 32 times and earned 82 citations, a rate of 2.56. Same site, same authority, same domain. The difference sits entirely in whether the page contains a sentence that answers a question on its own. We've unpacked the mechanics of being used versus being cited, and this is what it looks like at page level on a real site.

Why did our visibility stall while our share of voice climbed?

This is the part that would have made us quit if we'd only watched one number. Across four full tracking weeks, our visibility, meaning the share of AI answers that mention us at all, went from 19.6% to 18.7%. It went down. Anyone reading that column alone would conclude the publishing program failed.

Underneath it, everything else moved the right way. Our share of total mentions in the field climbed from 25.3% to 31.5%. Our average position improved from 2.3 to 1.9, meaning when engines name us we're now typically the first or second name out. Raw mentions went from 616 to 721 in a week. We're named more often, earlier, and in a larger share of what gets said, while the percentage of conversations that mention us at all held steady.

For context on the field: April Dunford's visibility over the same four weeks moved from 21.8% to 17.7%, and her average position from 2.2 to 2.7. That crossover isn't a verdict on her work, which is excellent and is why she's the reference point in this category at all. It's a measurement of publishing cadence against a fixed set of buyer questions. A book earns its authority once. A tracked prompt set asks the question again every morning. If you're reading a visibility dashboard this week, read it the way we've described in how to read AI visibility tools without being fooled, because a single flat number hides both the win and the loss.

The practical rule we now use: visibility tells you whether the category knows you exist, and it moves on a scale of months. Position and share of voice tell you whether this month's work landed, and they move in weeks. Watching visibility alone for the first six weeks of a publishing program is the fastest way to kill a program that's working. We nearly did it in week three.

What do the cited pages have in common?

We went back through every page above a 1.0 citation rate looking for the shared trait. It isn't length, schema markup, or publish date. Every one of them resolves a choice out loud. They say this option fits a company at this size with this problem, and this other option doesn't, and here's the line where the difference sits.

That's why the listicle wins, and the reason has nothing to do with the format. A comparison page can't be written at all until someone has decided who the company is for and who it isn't for. The format forces the decision. An essay lets you defer it forever while sounding thoughtful. When a buyer asks an engine who should rebuild their messaging, the engine is looking for fit statements, and it will quote whoever wrote one.

Which puts the real bottleneck somewhere most AEO advice never goes. If your company hasn't decided who it's for with enough specificity to sit in one row of a table next to a competitor, you can't write the page that gets cited. You'll write around it, publish something graceful, and get retrieved and dropped. That's the same failure that makes AI engines ignore your website wearing a different costume. Your narrative identity is the input. The citation is the output.

There's a test you can run on any draft before publishing it. Find the single sentence a stranger could lift and paste into an answer without needing the paragraph above it or the one below it. If that sentence doesn't exist, the page will get retrieved and dropped, no matter how well it reads. Our zero-citation pages all fail that test. Every page above a 2.0 citation rate passes it in the first screen.

What would we publish next if we started today?

In this order, based on what our own data rewarded:

1. 1One honest comparison page naming your real alternatives, including the ones you lose to. Put yourself in your true slot, not the top one.
2. 2One narrow vertical list for the segment where you actually win. Our healthtech page out-earns pages with triple the traffic.
3. 3A pricing or cost page that states real numbers. Ours quietly outperforms most of the blog on citation rate.
4. 4A homepage that passes the Cover-the-Logo Test, because it's the third-most-retrieved page on our site and engines read it as your definition of yourself.
5. 5Then the essays, written for the human who arrives after the engine has already named you.

You can see roughly where you stand before writing anything. Run your homepage through the Brand Signal Score and you'll get the 19-signal read on what an engine can extract from it today. If you'd rather watch this get built live, we run the AI Workforce Clinic weekly, and the first one is free.

One caution on all of the above. Five weeks is five weeks. We're reporting what our instrument recorded on our site in one competitive field, and we'll tell you when a pattern reverses. This is just truth, measured on the only company we can publish the raw numbers for.