The short version
- Before changing anything, read what the engines say across your buyers' questions. A spreadsheet is enough to start; one check proves nothing.
- Fix your own site first: it is what the engine reads when it searches, and it is the fastest thing a team can change.
- Then work on the places others talk about you: reviews, comparisons, communities and video.
- Drop what shows no effect on day one, and re-ask the same questions at the end.
What should you do before changing anything?
What to do. Start with a broad first reading. Put twenty or thirty of your buyers' questions to three engines, such as ChatGPT, Claude and Gemini, and note in a spreadsheet whether each answer names you and how it describes you.
Why it's here. Every later step depends on this reading: which questions leave you out, which criteria the engine uses, which pages it read. Be honest about what a hand check is: a first map, not a measurement. Single answers change each time a question is asked (What is GEO? has the numbers), so the figure that holds up is the share of answers that name you, taken across many questions and repeated readings, and read alongside the wording. Tools call it mention rate or share of voice; a spreadsheet is enough to start.
A fix chosen before reading the answers is a guess with a budget.
What it looks like. The marketing lead at a time-tracking tool for design agencies puts twenty-five buyer questions to three engines and logs the answers in a spreadsheet. Comparison answers mention the tool cautiously, as "a simple timer for freelancers". Questions about keeping projects on budget don't mention it.
Who usually owns it. Product marketing, since it owns the positioning.
Which questions should you track?
What to do. Track the questions your buyers actually ask, in their words, from the first problem through to the final choice.
Why it's here. The first reading shows which questions matter; now freeze them. Every later fix is measured against this set, so it has to stay the same from here on. Comparison questions deserve extra weight. They are where buyers choose, and they tend to send the engine to the web: in Graphite's research from August 2026, 58% of ChatGPT's answers to comparison prompts showed signs of a web search when it was left to decide. We think that matters: an answer built from a search is one your own pages can reach.
What it looks like. The team writes questions as agencies ask them on sales calls. "Why do our projects keep running over budget?" "Is a time tracker built for agencies worth it over a general one?" Neither names the product.
Who usually owns it. Product marketing, with the buyer's words from sales and support.
What should you fix on your own site first?
What to do. Rewrite the pages that answer your tracked questions, comparison questions first, so the answer, the criteria and the proof sit on the page itself.
Why it's here. The site comes first because it is what the engine reads when it searches, and it is the fastest thing a team can change.
Work through each page in this order. First, put the answer directly under a heading that matches the question. In an AirOps study from April 2026, ChatGPT cited question-matching headings 41% of the time, against 29% for the rest. The same study found that focus beat coverage: pages covering 26–50% of the sub-topics an engine searched for a question were cited more than pages covering all of them. In plain terms, one question answered well beats a page that tries to answer everything around it.
Second, state the criteria plainly (price, fit, who it's for), since the engine picks its own. Third, give it figures to quote: in the original GEO paper (KDD 2024), adding citations, quotations and statistics raised a page's visibility, though in a test engine rather than ChatGPT itself.
Last, update what exists rather than only publishing new pages. In Seer Interactive's July 2026 study, 72% of cited pages were fresh by last-updated date, against 42% by first-publish date, across four industries, none of them B2B software.
What it looks like. The pricing page gains a heading in the buyer's words, "How much does time tracking cost for a 20-person studio?", with the per-seat price in the sentence under it. The product page says the tool is built for agencies that bill by project, and that hourly freelancers may do better elsewhere.
Who usually owns it. Content writes, web ships, product marketing signs off the claims.
Where else do engines read about you, and what comes next?
What to do. Work on the places other people talk about you: review sites, comparison pages, communities and video.
Why it's here. This is where the evidence says much of the weight sits. In a December 2025 Ahrefs study of 75,000 brands, YouTube mentions tracked AI visibility most closely, branded web mentions close behind, backlinks far less. That is correlation, not proof of cause. It comes after the site because it is slower and mostly in other people's hands.
We also think a brand should publish its own honest comparison, so a competitor's is not the only one to quote. Why ChatGPT describes your brand the way it does shows what happens otherwise.
What it looks like. The answers drew on a review site, an old "best time trackers" list and a forum thread. The team asks agency customers for honest reviews, records a short video of the budget view, and publishes a comparison with general-purpose timers that says where each is the better pick.
Who usually owns it. PR and community for reviews, forums and video; product marketing for the comparison.
What should you stop spending time on?
What to do. Stop the work that has shown no effect, and the work that can only backfire.
- Schema markup as a citation play. In a May 2026 Ahrefs test of 1,885 pages, adding schema to pages AI engines already cited brought no measurable citation gain. Keep it for what it's for, clear facts about your product, but don't add it to get cited.
- llms.txt. It showed no relationship with citations across about 300,000 domains, in SE Ranking data reported by Search Engine Journal in November 2025.
- Checking one answer, or a "rank". In SparkToro's January 2026 study, the list of brands changed almost every time the same question was asked.
- Betting on one platform. Reddit's share of ChatGPT citations fell from 3.8% to 0.5% between 14 and 17 August 2026, in Promptwatch data reported by Semrush (provisional).
- Fake reviews and manufactured authority. We won't help with them, and you shouldn't try.
Why it's here. Stopping costs nothing, so do it on day one. It is listed here because the hours it frees only matter once the site and off-site work show where to spend them.
What it looks like. An llms.txt file and a schema sprint come off the backlog; the hours go to the pricing page.
Who usually owns it. Whoever owns the web backlog, usually web or SEO.
How will you know it worked?
What to do. Ask the same questions again and compare two things: the share of answers that name you, and what those answers say.
Why it's here. It closes the loop, and only works if the questions stay fixed. In SparkToro's January 2026 study, how often a given brand appeared across many runs was stable, even though single lists were not.
We have no evidence for how long a change takes to show, so we don't promise one.
What it looks like. The team re-asks its questions and checks how many budget answers now name the tool, and whether comparison answers still call it "a simple timer for freelancers". A change in wording counts, even when the share hasn't moved.
Who usually owns it. Whoever did the first reading, so both are read alike.
Your first month
Tools in this category, ours included, automate the first reading and the re-check. Ours also turns the site and off-site work into an ordered plan; How it works shows how.
Sources
- Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024 — source
- Graphite, "Do Not Force Search in Prompt Tracking", August 2026 — source
- AirOps with Kevin Indig, "The fan-out effect", April 2026 — source
- Seer Interactive, content recency and AI visibility, July 2026 (47,097 citations, four industries) — source
- Ahrefs, brand visibility correlations across 75,000 brands, December 2025 — source
- Ahrefs, schema markup and AI citations, 1,885 pages, May 2026 — source
- SE Ranking via Search Engine Journal, llms.txt across ~300,000 domains, November 2025 — source
- SparkToro with Gumshoe, AI recommendation consistency, 2,961 runs, January 2026 — source
- Semrush citing Promptwatch data, Reddit's share of ChatGPT citations, August 2026 (provisional) — source