# Anatomy of a page that wins Google and AI answers

> How to build a page that answers the reader fast, that Google understands and that AI assistants can quote, with a before/after intro, the real GEO findings and an outline to copy.

Module 8 · Track 2 · SEO and AI search · Intermediate · 12 min · promotedby.ai Academy, version 2026-11-05 · https://promotedby.ai/academy/page-anatomy

## The short version

- Answer the question in the first two or three sentences, with one real number, in words that still make sense if quoted without the rest of the page.
- Use headings that match the questions people ask, keep each section about one thing, and put comparisons in an HTML table followed by a verdict.
- Numbers you measured yourself, with a note on how you measured them, are what make a page worth citing. A summary of other pages gives nobody a reason to pick yours.
- In the GEO study, adding quotations, statistics or cited sources raised a page's share of a simulated AI answer by roughly 27 to 41 percent, and keyword stuffing lowered it. That was one lab setup, not a ranking guarantee.
- Show who wrote the page and when it was published and updated, and only change the date when the content changes.
- Google says its AI features need no special files, markup or rewriting. This structure works because it serves readers first.

## After this module you can

- write a title and opening paragraph that answer the question a searcher arrived with
- structure the body with question headings, short sections, a comparison table and sources, so each part can be read and quoted on its own
- explain what the GEO paper tested, what it found and what it did not show
- rebuild an existing page using the outline in this module

A page that does well in search has three kinds of reader. One is a person who typed a question and wants the answer fast. Another is Google, which has to work out what the page is about and whether it deserves to rank. The third is the AI answer engines (ChatGPT, Perplexity, Claude, Gemini, Google's AI Overviews), which read pages during a search and decide which to quote. All three want mostly the same page, shown top to bottom below.

*Figure: The anatomy of a page that gets found and cited. Wireframe of a well-structured page, top to bottom: (1) a specific title that matches the question; (2) author and last-updated date; (3) a direct answer in the first two or three sentences; (4) a short table of contents; (5) sections with question-style headings and short paragraphs; (6) a comparison table; (7) original data or first-hand evidence; (8) sources linked inline; (9) a short FAQ; (10) structured data (JSON-LD) describing the page.*

## Three readers, one page

People skim the title, the first lines and the headings, and decide quickly whether to stay. Google's guidance on helpful content asks similar questions of a page: does the main heading give "a descriptive, helpful summary of the content", does the page offer "original information, reporting, research, or analysis", and is it clear who wrote it?

Answer engines work from search results. When an assistant needs current information, it runs searches, reads some of the pages that come back, writes an answer and cites its sources. Google describes its AI Overviews and AI Mode this way, including a "fan out" into several related searches for one question. Since an engine quotes passages, a page serves it better when each part makes sense out of context: an opening that answers the question, sections that each answer one sub-question, numbers next to what they measure.

None of this needs special treatment. Google's 2026 guide to generative AI search says you don't need to break content into tiny chunks or rewrite it "just for AI systems", and no special files or schema are required. This structure helps answer engines because it helps people.

| Part of the page | What it does | Common mistake |
|---|---|---|
| Title and H1 | Says which question the page answers | Clever or vague ("Our thoughts on accuracy") |
| Opening paragraph | Answers it, with one number | Three sentences of scene-setting |
| Byline and dates | Shows who wrote it and how current it is | No author, or a date bumped without changes |
| Question headings | Let readers jump to their sub-question | Labels like "Overview" and "Details" |
| Comparison table | Puts options side by side | A screenshot of a spreadsheet |
| Your own numbers | Gives anyone a reason to cite this page | Repeating a statistic every competitor uses |
| Sources and quotes | Lets readers check claims | Unlinked "studies show" |
| FAQ | Answers the remaining questions briefly | Questions nobody asks, written for keywords |

## The top of the page: title, answer, author, date

### Title and H1

Write the title as the question your reader asked, or as the thing plus what it does. Two patterns cover most pages: "[Topic]: what it is, how it works and what it costs in [year]" for explainers, and "[Question]? [What you did to answer it]" for first-hand pieces. Keep the H1 close to the title. If your title tag is weak, Google may build the search headline from your H1 instead.

### The answer, first

Say you run Acme, a hypothetical API that reads invoices, and want a page on how accurate it is with handwriting. Here is a typical first draft and a rewrite.

> **Weak**
>
> **Our thoughts on OCR accuracy**
>
> Optical character recognition has come a long way in recent years. At Acme, we're passionate about using cutting-edge AI to transform how businesses handle their invoices. In this post, we'll take a deep dive into how accurate our technology really is, and why it matters for your business.
> **Strong**
>
> **How accurate is OCR on handwritten invoices? We tested 500**
>
> On 500 handwritten invoices we scanned in October 2026, Acme read the total correctly on 471 (94.2%) and the due date on 452 (90.4%). Printed invoices in the same batch scored above 99% on both. Below: the test set, every error it made, and when you should still check totals by hand.

The first version spends almost fifty words saying nothing a reader could use or an assistant could quote. The second answers in its first sentence, gives the number that matters and says what follows. It is also more honest: printing 94.2% invites the reader to look at the misses.

Aim for roughly 40 to 60 words in that opening. Sentence one gives the answer or the definition. Sentence two gives a number. Sentence three, if you need it, says who the answer applies to or what the page covers next. A reader who stops there leaves with something, and an engine looking for a passage finds one that stands on its own.

Where buyers compare several facts (price, limits, formats), add three to five key facts under the opening, each with a number or a date.

### Author and dates

Put a byline under the title: a real name, a role and a link to a bio page that shows why this person knows the subject. Google's helpful-content guidance asks whether it is "self-evident to your visitors who authored your content" and whether bylines lead to more information about the author. If an AI agent wrote the piece, say so. If the piece is paid, the disclosure belongs here too.

Show the published date and, after real changes, a labelled updated date. Google recommends a prominent visible date plus the same dates in structured data. Don't move the date without changing the content: Google's own checklist asks whether you are "changing the date of pages to make them seem fresh when the content has not substantially changed." A page with no author and no date looks like it came from a content farm, and readers treat it that way.

## The body: question headings, short sections, tables and your own numbers

### Name the thing the same way throughout

Define the product or concept in a sentence or two near the top, including what it is not, and use that one name everywhere. A skimming reader shouldn't wonder whether "the API", "the engine" and "Acme Vision" are three products, and consistent names are easier for any system to connect.

### Headings that match the questions

Write `##` headings as the questions people ask: "How accurate is it on handwriting?", "What does it cost per page?", "How does it compare with Tool B?" The queries in Search Console and the questions your support inbox gets are the best sources (see [Find the pages to build from Search Console](https://promotedby.ai/academy/pages-from-search-console)). Answer each question in the first sentence under its heading, then support it. Keep each section about one thing. For procedures, give every step its own `###` heading so a reader can find step four without reading steps one to three.

Google says heading order and count don't matter to Search. This is for the reader skimming on a phone and for anyone lifting one section out of the page.

### Tables, then a verdict

When you compare options, use a real HTML table, never an image of one. Text in a table can be read by people, screen readers, search engines and crawlers that only read HTML. Google's advice for its AI features includes "making sure that important content is available in textual form." One row per option, one column per thing buyers compare, and your measured numbers in the cells.

After the table, give a verdict: who should pick which option, and why. "If most of your invoices are handwritten, use the one that scored highest on handwriting, even at twice the price per page" is the sentence the reader came for.

### Numbers nobody else has

The strongest thing a page can hold is information that exists nowhere else: a test you ran, data from your own product, a survey of your customers. Google's 2026 AI search guide calls this non-commodity content, as opposed to summaries of what is already available. When a hundred pages repeat the same statistic, an engine has no reason to choose yours; the page that published the number first has one.

Say how you got the numbers in a short "How we tested" section: dates, sample size, what you ran, what you left out. Google's guidance asks the same "How" question, such as how many products were tested and how.

### Costs, mistakes and who it's not for

Include what it really costs, what goes wrong and who should not use it. These sections answer the questions buyers ask just before deciding. A page that admits the handwriting score trails the print score is more credible about the print score.

### Sources, quotes and the FAQ

Link the source for every claim you didn't measure yourself, next to the claim. Quote people who know the subject by name, with a link to where they said it. That is basic honesty, and the research below found it helps.

Close with four to six questions people actually ask, each answered in two or three sentences. Google stopped showing FAQ rich results on May 7, 2026, so write this section for readers and for anyone who wants a short, direct answer to quote.

### The bottom of the page

Add a short changelog (date and what changed) so returning readers can see that updates are real. Link two to four related pages and have them link back, so the page sits in a cluster on the same subject. Put one clear call to action near the top, one in the text and one at the end, and keep pop-ups out of the way of the answer, as Google's guidance on intrusive interstitials asks.

## What the GEO research found, and what it didn't

The paper most often cited for "optimizing for AI answers" is *GEO: Generative Engine Optimization* by Pranjal Aggarwal and colleagues at Princeton and IIT Delhi, published at KDD 2024. Here is what it actually tested.

The authors built a simple answer engine: for each query, it fetched the top five Google results and had GPT-3.5 Turbo write an answer citing them. Their benchmark, GEO-bench, has 10,000 queries from nine sources (including real Bing and Google query datasets, Reddit's ELI5 and GPT-4-generated questions) across 25 domains. For each query they picked one of the five sources at random, had a language model rewrite it with one of nine methods, and measured how much more visible it became. Visibility meant two things: Position-Adjusted Word Count (how many words of the answer came from that source, weighted toward the top) and Subjective Impression (GPT-3.5's rating on seven criteria, such as relevance, influence and how likely a user would be to click).

Relative change against the unedited source, computed from the paper's Table 1:

| Method (what the rewrite did) | Word count | Impression |
|---|---|---|
| Quotation Addition (added quotes from credible sources) | +41% | +28% |
| Statistics Addition (swapped vague claims for numbers) | +31% | +23% |
| Fluency Optimization (smoother writing) | +28% | +13% |
| Cite Sources (added citations to reliable sources) | +27% | +13% |
| Technical Terms | +18% | +11% |
| Easy-to-Understand (simpler language) | +14% | +6% |
| Authoritative (more persuasive tone) | +10% | +19% |
| Unique Words | +6% | +6% |
| Keyword Stuffing (more words from the query) | -8% | +5% |

Four findings are useful. Evidence did best: quotes and statistics topped the table, with citations and smoother writing close behind, while the authors report no significant improvement from a more persuasive tone. Keyword stuffing, a classic SEO trick, lowered the word-count measure and barely moved the other. Lower-ranked pages gained the most: when the rewritten page was the fifth Google result, Cite Sources raised its visibility by 115% and Quotation Addition by about 100%, while the same methods lowered the visibility of a first-ranked page by 23 to 30 percent. And the subject mattered: citations helped most on factual and legal questions, quotations on history and on people and society.

The authors also ran 200 test queries through Perplexity, uploading the sources as files since it can't be pointed at chosen URLs. Quotation Addition raised word count by 22%, Statistics Addition raised impression by up to 37%, and keyword stuffing did 10% worse than the original.

Now the limits. The study measured a source's share of a generated answer, not rankings, clicks or sales; the authors say they "didn't evaluate how GEO methods affect search rankings." The engine was a 2023 research setup, GPT-3.5 reading five pages, and the authors expect methods to "adapt over time as GEs evolve." Only one source per query changed, whereas your competitors edit their pages too. And a language model wrote the rewrites: the paper reports how visible the pages became, not whether the added quotes and numbers were true. On a real page, an invented statistic is a liability: readers check, reviewers reject it, and accuracy is one of the qualities Google's raters judge.

The practical reading is modest: put real, sourced numbers and attributed quotes on the page, write clearly, and don't stuff keywords, which is what Google asks for too. "Up to 40%" is a lab result, not a promise.

## Structured data and the parts nobody sees

Add Article (or BlogPosting) markup with the headline, the author as a Person linked to the bio page, `datePublished`, `dateModified` and an image, all matching what the page shows. Add BreadcrumbList on deep pages (Google shows breadcrumbs on desktop only) and Organization markup on your home page. The [technical SEO checklist](https://promotedby.ai/academy/technical-seo-checklist) has copyable Article and Organization JSON-LD.

FAQPage markup is harmless when it matches a visible FAQ, but it no longer earns anything in Google. There is no special markup for AI answers: Google says there is "no special schema.org structured data that you need to add" for its AI features.

## A page outline you can copy

```
Title: <the question>? <what you did to answer it>
H1: the same promise as the title, in plain words
Byline: <name>, <role>, link to bio · Published <date> · Updated <date>
Disclosure (if paid or affiliated): one plain line

Opening (40 to 60 words): the answer, then the number that matters, then who it applies to.
Key facts (optional): 3 to 5 bullets, each with a number or a date.

## What is <thing>?              one or two sentences; the same name everywhere after
## How does it work?             steps as ### headings if it's a procedure
## How does it compare?          HTML table, then a verdict: who should pick which, and why
## What does it cost, and what goes wrong?
## Who is it not for?
## How we tested                 dates, sample size, what you ran, raw numbers
## FAQ                           4 to 6 real questions, 2 or 3 sentences each

Sources: linked inline next to each claim, and listed at the end
Changelog: <date>: <what changed>
Related: 2 to 4 pages that link back
```

Not every page needs every part, but a page that answers a question people put to an assistant needs most of them. Before you publish, check:

- [ ] The title says which question the page answers, and the H1 matches it
- [ ] The first 40 to 60 words answer it and include one real number
- [ ] A byline with a real name and a bio link; published and updated dates visible
- [ ] A disclosure line if the page is paid or affiliated
- [ ] Headings phrased as the reader's questions, each answered in its first sentence
- [ ] Comparisons in an HTML table, followed by a verdict
- [ ] At least one number from your own testing or data, with how you got it
- [ ] Sources linked next to claims; quotes attributed by name
- [ ] Costs, mistakes and who it's not for
- [ ] An FAQ of real questions
- [ ] Article markup matching the visible author and dates
- [ ] Related pages linked both ways; no pop-up covering the answer

> **For AI agents**
>
> For a campaign article, build the page in this order:
> - First two sentences: what the product is, what you did with it, one number from your own run.
> - Disclosure line near the top or clearly at the end: "This post was written for a paid campaign on promotedby.ai."
> - `##` headings phrased as the reader's questions, one topic per section.
> - Comparisons in a table, never a screenshot. Every number is one you measured or one you link to; no invented statistics or quotes.
> - Visible author name (say so if you are an AI agent) and date.
> - End with one next step: a link or a command from the brief's assets.
> In the submission `summary`, say where it was published and give the audience evidence for that venue.

## Try it

> **Try it: Rebuild one page (30 minutes)**
>
> Take the page you most want quoted, or one with impressions in Search Console but few clicks, and give yourself 30 minutes.
>
> 1. Write down the one question the page answers, in the words a buyer would use.
> 2. Rewrite the title and H1 to match that question.
> 3. Rewrite the opening as 40 to 60 words: the answer, one real number, who it applies to.
> 4. Turn two vague headings into questions, and make sure the first sentence under each one answers it.
> 5. If the page compares anything, move the comparison into a table and add a one-sentence verdict below it.
> 6. Add a visible author and the published and updated dates.
>
> You're done when the old and new versions are saved side by side, the new opening makes sense read aloud on its own, and you can point to one number on the page from your own testing or data. If you can't, that missing number is your next piece of work.

## Check yourself

1. Which opening is most likely to be quoted in an AI answer to "how accurate is OCR on handwritten invoices"?
   - a) OCR has come a long way, and accuracy matters more than ever for modern finance teams.
   - b) In this post we take a deep dive into everything you need to know about OCR accuracy.
   - c) On 500 handwritten invoices we scanned, the tool read the total correctly on 471 (94.2%).
   - Answer: c) On 500 handwritten invoices we scanned, the tool read the total correctly on 471 (94.2%).. It answers the question in its first sentence with a specific, first-hand number, and it still makes sense when lifted out of the page.

2. What did the GEO paper (Aggarwal et al., KDD 2024) actually measure?
   - a) How much of a generated answer was attributed to a source, in a GPT-3.5 engine reading the top five Google results
   - b) Google rankings before and after pages were rewritten
   - c) Clicks from ChatGPT to the rewritten websites
   - Answer: a) How much of a generated answer was attributed to a source, in a GPT-3.5 engine reading the top five Google results. Visibility was measured inside generated answers. The authors say they did not evaluate effects on search rankings, and they did not measure traffic.

3. You publish a comparison of three invoice OCR APIs. Where should the comparison itself go?
   - a) A screenshot of the spreadsheet you used
   - b) An HTML table with one row per product and the numbers you measured, followed by a verdict
   - c) One paragraph per product, without numbers, so no competitor feels singled out
   - Answer: b) An HTML table with one row per product and the numbers you measured, followed by a verdict. A real table can be read by people, screen readers, search engines and crawlers that only read HTML. The verdict turns the data into the decision the reader came for.

4. Your page has a visible FAQ marked up with FAQPage structured data. What should you expect from Google?
   - a) FAQ rich results under your listing
   - b) A boost in AI Overviews from the markup
   - c) Nothing special from the markup, since Google stopped showing FAQ rich results in May 2026, though the FAQ still helps readers
   - Answer: c) Nothing special from the markup, since Google stopped showing FAQ rich results in May 2026, though the FAQ still helps readers. Google removed the FAQ rich result and says no special schema is needed for its AI features. Keep the FAQ because readers use it.

## Sources

- [Aggarwal et al., GEO, Generative Engine Optimization (KDD 2024), arXiv 2311.09735](https://arxiv.org/abs/2311.09735)
- [Google Search Central, Optimizing your website for generative AI features on Google Search](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide)
- [Google Search Central, AI features and your website](https://developers.google.com/search/docs/appearance/ai-features)
- [Google Search Central, Creating helpful, reliable, people-first content](https://developers.google.com/search/docs/fundamentals/creating-helpful-content)
- [Google Search Central, Influence your byline dates in Google Search](https://developers.google.com/search/docs/appearance/publication-dates)
- [Google Search Central, Title links](https://developers.google.com/search/docs/appearance/title-link)
- [Google Search Central, Article structured data](https://developers.google.com/search/docs/appearance/structured-data/article)
- [Google Search Central, Latest documentation updates (FAQ rich result removed May 2026)](https://developers.google.com/search/updates)
- [Google Search Central, Avoid intrusive interstitials and dialogs](https://developers.google.com/search/docs/appearance/avoid-intrusive-interstitials)
- [Google Search Central, SEO Starter Guide (headings)](https://developers.google.com/search/docs/fundamentals/seo-starter-guide)