Transcript
Intro
We took the opening paragraph of a film page and scored how well it matches the question people actually type: when is this film coming to streaming. The score was 52 out of 100. One rewrite later, the same paragraph scored 88. We added no research and no extra words, only the answer and the names. Let's see what changed.
The page belongs to ReelRadar, the made-up film site we use for every demo in this series, and the film is made up too. We'll rewrite it step by step, build a table of what the top results all forgot to say, and check what Google itself says about chopping content into little pieces for AI.
First, the measuring tool, since it comes back later in the series. An embedding model turns text into a list of numbers that captures what it's about. Two texts about the same thing end up with similar numbers, and the similarity score says how close they are. It's free, it's quick, and it's a fair stand-in for how many retrieval systems match passages to questions.
The answer comes first
Start with a habit that sounds obvious and almost nobody follows. Every question has an answer type. A when question wants a date, a where question wants a place or a platform, how long wants a duration, and who wants a name. If your first sentence doesn't contain that type of thing, it hasn't answered anything yet.
Here's the before. Get ready, film fans! Few films this year have generated quite as much buzz as this one, and everyone is talking about it. Now look for a date, a place, or even the name of the film, because there isn't one. The second paragraph offers an acclaimed director, a cast you'll recognise and glowing reactions, and still manages to name nobody.
And here's the after. Harbour of Glass opens in UK cinemas on 14 March. No streaming date has been announced yet, and the page says so, with a promise to add it the day the distributor confirms. That second sentence matters. For a where to watch question, not announced yet is a real answer, and it's the one most pages hide behind a paragraph of hype.
Now the scores. Against the streaming question, the opening paragraph went from 52 to 88. Against a release date question, it went from 53 to 92. The second paragraph rose too, from the fifties to the upper seventies, because it now names the film, the director and the actor. The facts were always available to the writer, and the only change is that the passage now carries them.
One warning before anyone builds a business on this number. The streaming question scored 88 against a paragraph that says there is no streaming date. Similarity measures whether a passage is about the right thing, not whether it's right. A model will happily match a confident wrong answer, so checking facts stays your job.
Why fuss over single passages at all? In October 2020, Google said it had started ranking with an understanding of specific passages, so a sentence buried deep in a page could still be found, and that this would improve 7 percent of search queries across all languages. AI answers pick sources passage by passage too. A clear passage gets found, and a vague one gets skipped.
You can test this on your own pages without any tools. Read only the first sentence under each heading. If that sentence answers the heading, the section works when it's quoted alone. If it's a warm-up, a definition nobody asked for, or a sentence starting with it, rewrite it until the answer comes first.
One idea per sentence
Second habit: name the thing before you use a pronoun. Inside a full article, the reader knows what it means. Lift a passage out on its own, as a snippet or as one source in an AI answer, and it points at nothing. So the first sentence of every section names the film, the person or the product it's about. After that, use pronouns like a normal person.
Keep one idea per sentence, and keep related words close together. If the release date sits three clauses away from the word cinemas, readers and parsers both have to work to connect them. Learn the verbs of your niche as well. In film, that's release, premiere, stream, direct and star, and each verb is a question people type.
Here's one sentence from the old page: out on the fourteenth in the UK it is, after a festival run where it went down well, with streaming to follow at some point. Three ideas, no film name, and a date with no month. Yoda would be proud. The rewrite splits it in two: Harbour of Glass opens in UK cinemas on 14 March, and no streaming date has been announced yet.
Headings in the order people ask
Third habit: order your headings the way people ask. The main question comes first, then the follow-ups in the order they arrive. For a film page, that's when and where it opens, then where to stream it, then who made it and who's in it, then how long it is and what it's rated. Each heading answers one question, and the first sentence under it gives the answer.
Where do the follow-up questions come from? Two free places. The people also ask box shows what Google thinks people ask next. The Search Console queries report for the page shows what people really typed before they landed. When a query keeps showing up with no matching section, that's your next heading.
There's a measured result behind this. In one controlled test, an ecommerce site rewrote the subheadings on its product pages as questions, so summary of features became what are the features of this product. The estimated result was 12 percent more organic traffic. One test on one site, so treat it as a strong hint, not a law.
What we wouldn't do is turn every heading on the site into a question by template. A heading should say what its section answers. Sometimes that's a question. Sometimes it's simply Cast, or Runtime and rating, and a reader scanning the page finds it faster that way. Nobody needs What is the runtime of Harbour of Glass as a heading.
Main content vs boilerplate
Fourth habit: keep the main content clearly apart from everything around it. Google's guidelines for its human quality raters define main content as any part of the page that directly helps the page achieve its purpose, and call it the reason the page exists. Raters don't rank pages, but the rubric tells you what Google is trying to reward.
Controlled tests point the same way. On one site, moving product descriptions out of tabs and up into view brought 14 percent more traffic from desktop users. On an ecommerce site, removing a product carousel from category pages brought an estimated 29 percent more organic traffic. Neither test added a single word. They just made the main content easier to reach.
On ReelRadar, that meant three changes. The answer now sits above the hero image instead of below it. The article lives inside a main element, so menus, newsletter boxes and promo carousels sit outside it. And the related films block moved to the bottom, where it helps people who finished reading instead of blocking people who just arrived.
Try this on your own site tonight. Open a page on your phone and scroll once. If you've met a hero image, a cookie banner, a newsletter box and an ad, but not the answer, your main content is losing a fight on its own page. The reader came for one thing, so let them find it before the furniture.
Information gain
Now the habit that separates pages that rank from pages that merely exist. Picture ten results that all say the same thing. Why would anyone, person or machine, prefer an eleventh copy? Google's own helpful content questions ask it straight out: does the content provide original information, reporting, research or analysis, and does it provide substantial value compared to other pages in search results?
Here's how we find that value on purpose. Take the top results for the query, here five pages about the Harbour of Glass release date, and pull every fact each page states into a table: one row per page, one column per fact. Then look at the empty cells. Every page gives the US date, and only one gives the UK date. None of them says whether a streaming date exists.
Those empty columns are the opportunity. The UK date, the age rating and a straight answer on streaming are things readers want, and the current results don't give them. You can do the extraction with a language model to save time, but check every cell by hand, because a wrong fact in your table becomes a wrong fact on your page. And only fill a gap you can actually verify.
What counts as a gain depends on the site. For a film site, it might be the UK date next to the US one, a runtime from the distributor's own listing, or a plain statement that no streaming deal has been announced. For a clinic, it might be real prices and recovery times from its own patients. Something the reader wants, that you can check, and that the other pages skipped.
You'll hear people call this an information gain score, as if it were a confirmed ranking factor. What actually exists is a Google patent with that idea in it, granted in June 2024, describing a score for how much new information a document adds beyond what a user has already seen. A patent tells you what Google invented, not what's running. Contested.
The verdict on adding something new that the results lack is simpler: it works. Google asks for it in plain words, and in a controlled test, pages with unique written content beat templated text by 14 percent. Just make sure what you add is something people want, not trivia bolted on to look different.
Two myths
Two myths sit right next to this topic. The first says longer content ranks better, so every page should hit some magic word count. We buried it in our evidence ladder episode, but it keeps climbing out, and rewrites like ours make pages shorter.
Google answers it in its own documentation, in the form of a question: are you writing to a particular word count because you've heard Google has a preferred word count? Then, in brackets: no, we don't. Our rewrite cut the opening from two paragraphs of hype to two sentences of facts, and the answer got easier to find. Placebo.
The second myth is newer. It says AI systems read in chunks, so you should cut everything into blocks of forty words or so, each one a self-contained answer, and rewrite your pages in a special style for AI.
In 2026, Google published a guide on its generative AI features with a list of things you don't need to do, and two lines settle this one. There's no requirement to break your content into tiny pieces for AI to better understand it. And you don't need to write in a specific way just for generative AI search. We couldn't find a forty word rule in any Google document. Placebo.
Look back at what actually changed on our page. We didn't cut it into forty word blocks. We put the answer first, named the film and the people, ordered the headings like the questions, and added what the other results lacked. That's ordinary good writing for people. It happens to parse well, which is exactly what Google's guide is telling you.
Wrap up
Here's the episode on one screen. Longer content ranks better: placebo. Chunking for AI: placebo, by Google's own account. Information gain as a confirmed ranking factor: contested, because all we have is a patent. Adding something new that the results lack: works, and it's the one habit here that also gives readers a reason to pick you.
If you do one thing this week, take one page that gets impressions but few clicks. Find its main query in Search Console, fix the first sentence so it answers that query, and fill the information gain table for it. Then wait a few weeks and compare clicks with the weeks before the change.
The page rewrite checklist and the information gain table template are free on the episode page at irongoo.com. Next time, we put the on-page advice you've heard over and over through actual tests, and see which of it survives.
