AI SEO7 min read
ByKarol PiaseckiFounder TextRank.ai
Published
Reading time7 min read
Updated

How to Write Content for AI SEO and GEO: Answer the Question, Not the Keyword

Learn what content AI systems pick for answers, what GEO and Ahrefs research shows, the mistakes teams make, and how long the work takes without tools.

A marketing team publishes “The Best Customer Onboarding Software.” It runs 2,500 words and repeats the keyword in the title, the headings and every third paragraph. A buyer asks ChatGPT for onboarding software for a 20-person SaaS team that integrates with HubSpot. The model cites a competitor’s page instead. That page is 900 words long, but it lists the user limit on every plan, explains the HubSpot sync and states how long setup takes.

That gap is what AI SEO and GEO (generative engine optimization) are really about. A language model produces the most probable answer to a question. When it searches the web, “most probable” in practice means the answer best supported by sources: with a condition, a number, an example and a limitation. Generic copy can’t supply that, even when it is popular and ranks well.

What research says about the content AI picks

The most cited study on this was published in 2023 by researchers from Princeton and collaborators. “GEO: Generative Engine Optimization” was accepted at KDD 2024. The authors built a benchmark of 10,000 queries across 25 domains and tested nine ways of rewriting source content.

Adding quotations, statistics and citations to sources performed best, improving a source’s visibility in the generated answer by up to roughly 40%. Keyword stuffing did little or nothing; on Perplexity it scored about 10% below the unmodified baseline. One result stands out for smaller sites: citing sources raised visibility of the page in fifth position by more than 100%.

The study has limits worth knowing before you turn it into a recipe. It used a custom engine built on GPT-3.5 with the top five Google results, plus a sample on Perplexity. Today’s ChatGPT, Gemini and Claude work differently. The direction still holds: models are more likely to use a passage that can be checked.

A second signal comes from Ahrefs data on Google AI Overviews. In July 2025, 76% of cited pages ranked in the top 10 for the same query. In a March 2026 update covering 863,000 keywords, the figure was about 38%, and 31% of cited pages ranked outside the top 100. Ahrefs notes that it changed how it detects citations, so the numbers are not directly comparable. It points to query fan-out, which Google describes in its documentation as running related searches across subtopics.

For an editorial team, this means ranking for the head term helps but isn’t enough. A page can be cited for answering a sub-question, such as the user limit or setup time. If that answer isn’t there, there is nothing to cite.

The May 2024 AI Overviews episode, when the feature suggested adding glue to pizza sauce, showed the same mechanism from the other side. Google explained that for unusual queries there was little high-quality content, so the system drew on satire and joke forum posts. When nobody answers specifically, anyone will.

How to build a passage AI can use

The unit of work is not the page. It is the passage: a paragraph, a table row or an FAQ answer that still makes sense when lifted out of context. A strong passage has five parts:

  1. The answer first. The opening sentence settles the question: “Yes, the HubSpot integration syncs both ways.”
  2. A named subject. Write “[Product]’s HubSpot integration,” not “our integration.” A passage pulled from your page still has to say who it is about.
  3. The condition. Plan, limit, region, version: “available on the Business plan, up to 50 users.”
  4. Evidence. A number, a date, a documentation link, a method, or a named customer if you have permission.
  5. A limitation. What the product doesn’t do, or when it isn’t the best fit. This adds credibility.

Compare two versions of the same paragraph. Before: “Our platform integrates easily with popular CRMs and saves your team time.” After: “[Product]’s HubSpot integration syncs contacts and companies both ways every 15 minutes. It is available on the Business plan. Custom fields need manual mapping during setup.” The second version answers three sub-questions. The first answers none.

Formats that often help: a one-sentence definition, a comparison table with explicit criteria, a short FAQ answer, a list of setup steps with durations. None of them is magic. They help because they force you to be specific.

The most common problems

Writing for the keyword instead of the question. The article repeats “best onboarding software” but never says who it is for, what it costs or which integrations it supports.

Facts scattered across pages. Pricing sits on the pricing page, limits in the docs, integration conditions in a sales PDF. No single passage holds the whole answer.

Marketing filler. “Comprehensive,” “intuitive,” “scalable.” The model has nothing to extract, and the reader has nothing to verify.

Contradictions between materials. The site says setup takes 14 days; the docs say “about a month.” The model may pick either version, or an outside source. See what to do when AI gets facts about your company wrong.

Faith in technical shortcuts. Google states plainly that AI Overviews and AI Mode need no new files, AI text files or special markup. For ChatGPT, robots.txt must not block OAI-SearchBot. These are access requirements, not a content advantage.

No retest. The team ships a change and never re-runs the same questions in the same models a few weeks later.

How long it takes without professional tools

Say you want to prepare one product page for five buyer questions and check them in four models: ChatGPT, Gemini, Claude and Grok. Our working estimate for manual work looks like this:

StageWhat it involvesEstimated time
Collecting answers5 questions × 4 models, saving answers and links1.5–2.5 h
Reading sourcesUsually a dozen to several dozen unique pages4–8 h
Comparing with your pageWhat sources have and you lack, passage by passage2–4 h
Gathering factsConversations with product, support and sales2–4 h
Editing and approvalRewriting passages, expert review3–5 h

That adds up to somewhere between the mid-teens and over twenty hours for one page and five questions. For ten pages, it is several weeks of one person’s time, and model answers will have shifted along the way. The expensive part isn’t writing. It is the diagnosis: working out what exactly is missing and why.

How TextRank shortens the work

TextRank handles the diagnosis. For a chosen buyer question, it collects answers from ChatGPT, Gemini, Claude and Grok, along with the sources each model cites. It then compares them with your page passage by passage, checking how directly each passage answers, how clearly entities are named, how dense the facts are and whether the passage stands on its own.

The output is a change plan rather than a single score: what to keep, what to rewrite, what evidence to add and whether you need a separate piece. Each recommendation links back to the prompt, the model answer and the source, so your product expert knows which fact to confirm. Collecting answers, reading sources and comparing, the biggest part of the table above, happen in the tool. Your team keeps the decisions, the fact-checking and the writing.

TextRank won’t invent facts your company doesn’t have, and it doesn’t guarantee citations. It shows what is missing and lets you check the effect in the next audit. To see how to analyze the cited sources themselves, read how to analyze the content ChatGPT cites. For B2B SaaS product pages, see our guide to GEO for B2B SaaS.

Audit your first page in TextRank.

Choosing analysis and briefing tools

Choose the question and establish the facts before drafting. A tool may help review answer completeness, topic context, or AI sources. The resulting brief should identify the missing information and its factual approver. A higher score does not replace that reasoning.