Iron Goo
P3E06 · Part 3: AI SEO

AI topical maps are guesses until you check search demand

Learn how to turn AI topical maps into content plans backed by real search demand. See why entity checks and query clusters need human review before you publish or merge pages.

What this episode covers

  • Drafting entity tables with clear attribute definitions, sources and review status.
  • Checking AI topic suggestions against Search Console and Google autocomplete.
  • Reading search results to confirm intent and choose the right page type.
  • Using your own customer questions to avoid generic content plans.
  • Clustering real queries to investigate content gaps and possible keyword cannibalisation.
  • Spotting multilingual pages that clustering tools wrongly flag as competing.

Key moments

  1. 00:00Intro
  2. 00:46Entities: draft, then check
  3. 04:06The 59-topic map
  4. 07:17Why everyone's AI map looks the same
  5. 08:35Query clusters with embeddings
  6. 09:49Clusters vs pages
  7. 13:00Wrap up

Ready to move?

Tell us where your business is today. A written assessment lands in your inbox within two business days.

Talk to us

Transcript

Intro

We asked Claude for a topical map for a dental clinic that treats patients from abroad. Fifty nine topics, seven tidy clusters, under a minute. Then we checked every topic against real search data. Thirty three of them didn't make it.

The AI didn't fail. It did exactly what a model does: it gave us the most likely map for a dental clinic. Tidy, confident, and roughly half wishful thinking. The skill is knowing which half.

So today AI drafts three things for us, an entity table, a topical map and query clusters, and then real data gets the red pen, on camera, with every number from our own runs.

Entities: draft, then check

Start with entities, the things your topic is made of. For an article on implant brands, we asked Claude for fifteen manufacturers, each with a type, a home country and a founding year. No web search. It came back in seconds and looked like an encyclopedia page. That's exactly when you check it.

Why entities first? Search engines work with things and their attributes, not just strings of words. The entity table is the skeleton of your map. Get one bone wrong, and every page built on it limps the same way.

So we wrote a checker. It looks up each brand in Wikidata, the open knowledge base, and compares type, country and year. Google's Knowledge Graph Search API does a similar job if you have a key. We expected to catch the AI out. The checker caught itself first.

Five of the fifteen lookups matched the wrong thing. Straumann came back as a family name. Two brands came back as scholarly articles. Neodent came back as a genus of algae, and Bicon as a genus of insects. Two more weren't found at all. A naive checker is exactly as confident as a naive writer.

So the checker got a rule of its own: the match has to be a company. Second run, nine brands matched the right company, nothing matched an insect, and six weren't in the reference at all. Those six go to a person with each company's own website open. That's the honest output of an entity check: a short list for a human.

Now the founding years. Of the eight real matches in the first run, four agreed, three disagreed and one was missing. Take Camlog. Claude said Germany, 1999. Wikidata said Basel, 2004. So we opened the company's own history page. The business started in 1988, the brand launched in 1999, and the Basel company was set up in 2004. Everyone was right, and nobody was useful.

So the real bug wasn't a wrong year. It was an undefined attribute. Founded can mean the brand, the company or the workshop it grew out of, and the AI quietly picked one. Henry Schein was a different miss: the table called it a manufacturer, and Wikidata calls it a distributor. Define every attribute before you generate, and let a person settle each disagreement from the primary source.

And treat the reference as a second opinion, not a judge. Anyone can edit Wikidata. When two sources disagree, the company's own page decides, and your table records which source you used.

Here's what a table that survives looks like. Every attribute gets a written definition, so founded means one thing for every row. Every value carries its source and the date someone checked it. And a status column says draft, checked or disputed. It's duller than the AI's version, and every row of it is true.

Claim one. The AI knows your niche's entities. It rattles off the famous names, so it sounds like an expert.

It knows the popular names. Then it fills the attributes with plausible values and never tells you what they mean. In our table, three of the eight companies we could match had a disputed founding year, and one was the wrong kind of company. Draft the table with AI, but never let it be the source. Verdict: placebo.

The 59-topic map

Now the map. We asked for exactly sixty article topics for the clinic's English blog, each written the way a patient would type it. We got fifty nine. Tiny thing. But it's the habit that matters: count the output before you trust anything inside it.

Every topic then faced two questions. First, does the clinic's own Search Console show a query with nearly the same meaning and real impressions? We matched them with Voyage embeddings, the provider Anthropic's docs point to, since Anthropic doesn't make its own. Second, does Google's autocomplete extend the phrase, a hint of demand somewhere else?

Twenty topics had clear demand in the clinic's own data. Eight had none there, but autocomplete said people search for them. Thirty one had no evidence anywhere. Watch the map shrink.

Then a person read the twenty green ones, and two were wrong. Istanbul dental trip itinerary had matched a French query about dental tourism in Istanbul. Turkey visa for dental treatment had matched dental treatment in Turkey. Close in meaning, different intent entirely. Nobody asking about a visa wants a price list. The script was confident and wrong twice, which is why a person reads every row it keeps.

Final count: eighteen topics with demand in the clinic's own data, eight with outside demand only, thirty three cut. Whole clusters vanished. All eight travel topics, gone. Five of the six smile makeover topics, gone. What survived looks like the clinic's real business: implants, veneers, crowns and what they cost.

Credit where it's due. Eleven of those eighteen survivors matched a real search query word for word, like dental implant cost turkey and how long do veneers last. Models are good at the obvious head topics of a niche. The trouble is everything they invent around them.

The eight amber topics get a different status. All on six implants and same day implants had outside demand, but nothing yet in the clinic's data. Treat them as bets. Build one or two, give them a few months, and let Search Console tell you whether the clinic can win those searches.

Keep the map in a file, not in a chat. One row per topic, with its evidence, its status and the date you checked. Re-run the same script every month. A grey topic that starts showing impressions turns amber. An amber one that earns traffic turns green. Now the map is a record, not a guess somebody made once in spring.

Two limits to know. Search Console only shows searches you already appear for, so a new site leans harder on outside evidence, and autocomplete alone is weak evidence. And before you build anything, look at the live results page. If Google shows clinic price pages for a topic you planned as a blog post, the page type is wrong, and no map will warn you.

And no evidence doesn't mean worthless. What to pack for a dental trip will probably never earn a search click. A patient who has already booked would love it in a confirmation email. Cut it from the search map and hand it to the team that looks after patients.

Why everyone's AI map looks the same

Now the part that should worry you. We sent the identical request twice more. In the second map, thirty two of sixty topics were near duplicates of the first. In the third, thirty seven of sixty two. Same prompt, mostly the same plan.

That's one model talking to itself. Now picture every clinic in Istanbul asking a model the same question this year. Models return the most likely answer, so the maps converge, and the same maps produce the same pages. Congratulations: you and your competitors now share a content strategy.

The way out is data only you have. Your Search Console queries, yes. Also the questions patients ask on the phone, the objections in sales calls, and the thing support explains every single week. Put those in the brief that drafts the map, and it stops being the average clinic's map.

Claim two. An AI topical map gives you topical authority. Generate the map, publish every node, and Google crowns you the expert.

A map is a plan. Authority comes from coverage that searchers and other sites actually use. Our check cut more than half the plan for lack of evidence, and what survived was the clinic's real services. Publishing all fifty nine would have meant thirty three pages written for nobody. Verdict: placebo.

Query clusters with embeddings

The third draft starts from the other end: the searches you already get. We took every query with at least ten impressions in the clinic's last four weeks. One thousand five hundred and twenty of them, in a lot of languages. Each became an embedding, a list of numbers that puts similar meanings close together. Then a clustering step grouped the neighbours.

Notice who does what. Embeddings do the grouping, cheaply and the same way every time. Claude does the human part: reading each cluster and giving it a name. Nobody asked a chat model to invent the groups, so nothing in them is made up. Every query was typed by a real person.

For naming, give Claude the five biggest queries in each cluster and ask for a short name and the searcher's intent. Nothing else. Then skim the names against the queries. A name that doesn't fit its queries is the fastest way to spot a cluster glued together wrongly.

One warning before you trust any of it. The number of clusters is a dial you set. Tighten it and the same queries split into six hundred and fourteen groups. Loosen it and they fall into one hundred and seventy. Turn the dial and your list of problems changes too. So borderline cases are questions, not answers.

Clusters vs pages

Now join the clusters to the pages Google shows for them. Two pages splitting one cluster might be cannibalisation. A cluster with no page of its own is a gap. One page spread across unrelated clusters is unfocused. On a multilingual site, every one of those comes with a twist.

At our setting, thirty six clusters had at least three hundred impressions. Eight of them were split, with two or more pages each taking at least a fifth of the impressions. Clicks were too few to judge, so we used impressions. On paper: eight fights to break up.

Then we read them. Five of the eight were the same article in different languages. The embeddings are multilingual, so Dutch and Spanish searches about sensitive teeth landed in one cluster, and each language version took its own share. That's not a fight. That's the site working exactly as designed.

The other three went on a short list for a person, with the results page open, because that's where a flag belongs. Had we merged pages straight off the flag list, we'd have flattened five language versions that were doing their job. That's an expensive way to learn what an embedding is.

So what does a real fight look like? Picture ReelRadar, our fictional film site, with a where-to-watch page and a release-date page for the same film, and both catching searches like is it streaming yet. One question, two half-answers. One page should own it, and the other should link to it.

Fixing a real fight is boring on purpose. Let the results page pick the owner. Move the half-answer from the other page into the owner, link across with a descriptive anchor, and check the cluster again in a month. If the split shrinks, you're done. If it doesn't, the two pages answer the question too similarly, and one of them needs a new job.

The unfocused check needed a human read too. The page spread across the most clusters was a guide to replacing missing teeth, split over four. We read them, and all four were missing tooth options plus a word like cost, guarantee or Turkey. The clustering had sliced one topic too thin. The page was fine, and the flag was just one more question for a person.

Gaps are the third pattern: a cluster with real impressions and no page that clearly owns it, so its searches scatter across pages written for something else. Same routine. Read the queries, look at the results page, then decide: a new page, or a better answer on an existing one.

For every borderline case, the tie breaker is the results page. If Google shows mostly the same results for two searches, it treats them as one intent, and one page should own both. If the results differ, two pages can be right. Embeddings are the fast first pass. The results page is the ground truth.

Claim three. Embedding clusters are your cannibalisation list. Run the notebook, get the list, merge the pages, go to lunch.

They're a list of questions. Five of our eight flags were language versions that should stay exactly as they are, and the number of flags moved with a dial we chose. Merging off that list would have damaged a healthy multilingual site. Verdict: contested. Great for finding candidates, never for deciding.

Wrap up

The whole episode on one screen. AI drafts are fast and useful, as long as something real decides. Entities get checked against a reference, and a person settles the disagreements. Map topics need evidence of demand. Clusters come from real queries, and a person reads every flag before anything gets merged.

If you keep one thing, keep the order. Draft with AI. Check against a reference or real demand. Have a person read what the check keeps and what it flags. Then build, and not a minute before.

What will change: the models, the embedding providers, and what their maps look like. What won't: a draft is a hypothesis, demand is evidence, and a person reads what the script keeps. The map validation script and the clustering notebook are free on the episode page at irongoo.com. Next episode: turning a validated map into briefs that produce publishable drafts.

Free ebook

Put your next steps in order

Build on what you have just read with a free ebook on semantic SEO and AI visibility. Get worked examples, decision rules, checklists and a do-this-week action in every chapter, plus a six-week plan for your site.

By signing up and confirming your address, you consent to receive the free ebook plus one email every two weeks with 3 or 4 new Iron Goo videos, guides and articles, and nothing else; unsubscribe any time with one click.