Transcript
Intro
Type the word Dune into a public knowledge base and eleven of the first fifteen results are called exactly that. A 1965 novel, films from 1984 and 2021, a video game, an album, and a hill of sand. Now imagine you're a search engine, and a page just said Dune at you. Which one did it mean?
This episode is about how machines make that call, how to make it easy for Google and for AI assistants alike, and which popular entity tactics are a waste of an afternoon.
Things, not strings
Google said it out loud in May 2012, when it launched the Knowledge Graph: things, not strings. Its example was Taj Mahal, which could mean a monument or a musician, and getting that wrong sends tourists to a blues gig. At launch the graph held more than three and a half billion facts. By 2020, Google said, it held over five hundred billion facts about five billion entities.
And the graph isn't only there to decorate the panel on the right of the results. Notes from a Google engineer, filed in the US antitrust case, say the Knowledge Graph is used beyond that panel. Ask about a famous person's spouse, and the graph tells regular search who that is, so the blue links end up about the right person.
Recognise, resolve, attach
We teach this as three steps. It's our working model, but each step shows up in Google's own public tools. Recognise: spot that a bit of text names a thing. Resolve: decide which real thing it is. Attach: pull in what's already known about it, and read what your page adds.
Resolution is where most pages lose. Here's that Dune list, pulled on the nineteenth of October 2026. Frank Herbert's 1965 novel, David Lynch's 1984 film, Denis Villeneuve's 2021 film, a film series, a franchise, a video game. If your page only ever says Dune, the engine is left guessing from whatever crumbs you dropped.
Sequels make it worse. Dune Part Two came out in 2024. Dune Part Three is scheduled for US cinemas on the eighteenth of December 2026, and it adapts the second novel, Dune Messiah. A page that keeps saying the new Dune movie will be describing a different film within a couple of years. Pages age, and nicknames age faster.
Once the thing is resolved, attributes pin it down: director, cast, release date, runtime, where to watch. These are the facts that fit one film and none of its namesakes. Name the film, its director and its source novel in your opening lines, and you leave very little room for doubt.
Superman has the same problem. Wikidata lists five different films with exactly that title, from 1978 to 2025. The title alone is a string. The title plus a year plus a director is a thing.
Queries help with resolution too. Someone typing Dune cast almost certainly means a film, and someone typing Dune author means the novel. Your page has no query to lean on, so it has to carry its own clues: what type of thing it is, the year, and the people attached to it. Give those clues before the engine has to guess.
Why repetition stopped working
Old keyword habits assume that saying a phrase more often makes a page more relevant. Modern retrieval doesn't work by counting. The court record describes RankEmbed as a deep learning system with strong language understanding that can find the best documents even when the query lacks certain words. Matching meaning beats matching letters.
Google has been heading this way for years. In 2019 it said its BERT models helped it understand one in ten English searches in the US, especially longer ones where small words like for and to change the meaning. A keyword counter can't see those words. A reader can't miss them.
So repetition buys you nothing once the thing is clear, and it costs you readability. Twelve mentions of Dune don't tell an engine whether you mean the book or the film. One sentence with the year and the director does.
Quick quiz. Which tells a machine more: Dune, Dune, the Dune movie everyone loves, Dune? Or: the 2021 Dune, directed by Denis Villeneuve? The second has fewer Dunes and far more information, and search engines grade it the same way you just did.
So where does the full name go? In the spots people read first and machines lean on most: the title, the main heading and the first sentence. After that, write like a person. Restate the full name when a new section starts and the subject could be in doubt, not in every paragraph.
Name it fully and early
Here's a check you can run yourself. Google Cloud's Natural Language API has a free demo that finds the entities in a paragraph and gives each one a salience score from zero to one, where higher means more central to the text. It's a Cloud product, not Google Search's ranking system. It does show you how one serious machine reader handles your writing.
Version one is a typical intro on our example site, ReelRadar. It gushes about the next chapter, talks about the sequel, calls the film it, and names it once, right at the very end. Very exciting, and very vague. We paste it in and see which entity the tool puts on top.
One line from the documentation for this API deserves a spot on screen. It counts proper names and common nouns as mentions, and pronouns aren't on its list of mention types. So every time the paragraph says it, the tool doesn't record a mention of the film at all.
Version two names the entity in the first sentence. Dune Part Three, Denis Villeneuve's film of Frank Herbert's Dune Messiah, is scheduled to open in US cinemas on the eighteenth of December 2026. Then it gets on with the trailer. Same facts, new order, and now we compare which entity comes out on top, and whether the film gets linked to a Knowledge Graph ID.
Whatever the scores say on the day, the rewrite follows four rules we use on every page. The full name, not a nickname. A disambiguator, like a year, a maker or a type. The answer the query wants. And pronouns only after the name has done its job.
You can run this on your own site in ten minutes. Open your ten most important pages and read only the first two sentences of each. If you can't tell exactly which thing a page is about, neither can a machine reading it in pieces. In our audits, that one check finds more problems than any entity tool.
Your business is an entity too, and the same rules apply. Name it the same way everywhere, say what type of business it is and where it operates, on your about page, your contact page and your profiles elsewhere. A clinic that calls itself one name on the homepage and a slightly different one on its listings is asking machines to guess whether those are the same place.
Two tactics on trial
Now for the tactic this idea usually turns into. Mention as many entities as you can, the advice goes, and Google will see how relevant you are.
This is keyword stuffing with extra steps. A paragraph listing every actor, planet and studio is a list of strings in a new costume. Google's spam policy describes keyword stuffing as filling a page with keywords to manipulate rankings, and gives blocks of listed cities and regions as an example. Swap the cities for actors and you've reinvented it. Placebo.
What works is relevance by relationship. Mention another entity when it explains yours: the director, because people search by director, the novel, because the film adapts it, and the earlier film, because the story picks up from it. Every mention should answer a question a real reader has.
Then be consistent across the site. Call the film by the same full name everywhere, and use that name as the anchor when you link to its page. Google's link guidance says good anchor text is descriptive and relevant to the page it links to, and that the words around the link matter too. Five nicknames for one film are five chances to be misunderstood.
Claim two. Entity SEO means adding schema markup. Put the right structured data on the page and Google knows exactly what it's about.
Structured data is useful, and Google says it uses it to understand a page. It also says not to mark up anything users can't see, and that AI Overviews and AI Mode need no special schema. Its 2026 guide to generative AI search even lists over-focusing on structured data as a misconception. Markup describes the entity your text already names, and it can't rescue a page that never says which Dune it means. Placebo.
So keep the markup, and give it the right job. It restates the facts on the page in a form machines parse easily, and it can win rich results where your page type qualifies. It doesn't replace the sentence that names the thing.
How AI assistants read it
AI assistants face the same problem, with one extra twist. Google says its AI features may fan a question out into several related searches across subtopics and sources. Other assistants retrieve in their own ways, but most of them work from pieces of your page rather than the whole thing.
That's where pronouns really hurt. Lift a passage out of the middle of your page that says it lands in December, and it has lost its subject. A passage that says Dune Part Three is scheduled for US cinemas on the eighteenth of December 2026 stands on its own. In our AI visibility work, passages that stand alone are the ones we see quoted.
So one rule serves both audiences. Name the thing in a way that survives being cut out of the page, and restate the full name at the start of each major section. Google's own 2026 guide says optimising for generative AI search is still SEO, and on this point it's right.
You can test the assistant side for free. Paste one section of your page, on its own, into an AI assistant and ask what it's about. If it names the right film, the right year and the right maker, the section stands alone. If it hedges or picks a namesake, you've found your next rewrite.
This works the same way far from film. A product page that calls the item this model, or a clinic page that keeps saying the procedure, has the Dune problem. The fix is identical: the full name, one disambiguator, then the details.
Wrap up
Here's the episode on one screen. Engines and assistants turn words into things, and pages lose at the resolution step. Listing more entities is stuffing. Schema restates what your text already says. Naming the entity fully and early is the move that works.
The entity clarity checklist is free on the episode page at irongoo.com, ready to run against any page you own. Next time, we take one entity and turn it into a full attribute plan, built from real search data instead of guesses.
