A 90-day GEO plan for a company brand, starting from zero

This plan is for the marketer or operations lead who has been told to “do GEO” for the company brand, with no earlier work to build on. It turns the first 90 days into a sequence: find out where you stand, remove technical blocks, fix the pages that matter, get described accurately by third parties, and set up measurement. Budget about half a day a week for one person in month one, rising to two or three days a week from month two, plus help from whoever runs the website.

The short version

1. Weeks 1–2: write down what you want AI to say about you, and the 20–30 questions customers actually ask.
2. Weeks 2–4: run a baseline across the assistants your customers use, with enough repeats to mean something.
3. Weeks 3–5: confirm search engines and AI crawlers can reach and quote your key pages.
4. Weeks 5–9: rework three to five core pages so each answers one real question clearly.
5. Weeks 6–12: get listed, correctly, on the list and comparison sites that answers in your category cite.
6. Weeks 10–13: re-measure with the same prompts, and report foundations built rather than promised rankings.

The phases overlap on purpose; the sections below follow the same order.

Set expectations before you start

GEO from zero is slow because much of the material AI draws on is outside your site. Two industry frameworks agree that the first quarter is for diagnosis and foundations.

DATA17’s 2026 action guide proposes a 90-day phase: month one sets a one-line positioning, a list of 20–30 questions and a baseline measurement. Month two maps the sources AI cites and fixes technical issues. Month three submits the brand to its top ten lists and directories, reworks three core pages and produces one asset for third parties. It warns not to expect the mention rate to rise noticeably within 90 days. This is the authors’ recommendation, not a tested result.

Hypothesis · DATA17 AI Research Lab · 解数智能 (unpublished report, not publicly available; readers can't check it)

Yisou Baiying’s 2026 evaluation white paper sets out a similar roadmap: a baseline in the first 30 days, a metrics system by day 90, and an ROI review only at day 180. The paper presents this as a management framework and reports no data showing it works.

Hypothesis · 一搜百应 GEO 研究院 (unpublished report, not publicly available; readers can't check it)

Tell your manager this at the start. A plan sold as “AI will recommend us in a month” will look like failure at week twelve.

The first 90 days buy you a map and a clean foundation. Mentions come later, and saying so up front is part of the plan.

Weeks 1–2: define the target and the questions

Before measuring anything, agree on two things.

• One sentence about what you are. Write a narrow, factual description: what you sell, to whom, and what makes it different. Get product and sales to sign it. Every page and listing later will reuse it.
• A question list. Collect 20–30 questions from sales calls, support tickets and search queries. Group them by stage: category (“best X for Y”), comparison (“A vs B”), problem (“how do I fix Z”) and brand (“is [brand] any good”).

Category and problem questions matter most, because they are how new customers find you. Brand-name questions nearly always return your brand and prove little (Measuring AI visibility by sampling prompts).

Weeks 2–4: run a baseline

Ask each question in the assistants your customers use and record whether you appear, in what position, which sources are cited, and whether the description is right.

Observed in 2025-09, more than 9 months ago. Answer engines change often; re-test before relying on this.

In a September 2025 study (2,025 runs: five typical questions, 135 runs each on three Chinese models), Miaozhen concluded that results reach basic stability only at 30 or more runs per prompt. Brands with low visibility need 60 or more.

Observed in practice · observed 2025-09 · 中国商务广告协会AI营销应用工作委员会 · 秒针营销科学院

That is more work than most teams expect. If you can’t run 30 repeats for every question, run fewer questions properly rather than all of them once. Keep the raw answers. The full protocol is in Measuring AI visibility by sampling prompts. Fix the question list now and don’t edit it later, or the next round won’t compare.

Weeks 3–5: confirm AI search can reach you

No content work helps a page that AI search can’t fetch. Ask your web team to run the AI search access audit on your homepage, product pages and main guides.

To appear as a supporting link in AI Overviews or AI Mode, a page “must be indexed and eligible to be shown in Google Search with a snippet”.

Officially documented · Google Search Central

OpenAI: “Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links.”

Officially documented · OpenAI

Checklist for this step:

• Key pages indexed in Google and Bing.
• robots.txt allows the search crawlers you care about (Configure robots.txt for AI crawlers).
• CDN or firewall doesn’t block them.
• No stray nosnippet or noindex (Snippet controls for AI Overviews and AI Mode).
• Main text present in the HTML, not only after JavaScript runs (JavaScript rendering and AI crawlers).
• For Chinese assistants, the equivalent checks on Baidu and others (GEO in China: Doubao, Yuanbao, Qwen, ERNIE, DeepSeek and Kimi).

Weeks 5–9: rework three to five core pages

Pick the pages that should answer your most important category and comparison questions. Rewrite them rather than creating dozens of new ones.

Google says “Creating content that people find unique, compelling, and useful will likely influence your website’s presence in generative AI search in the long run more than any of the other suggestions in this guide.”

Officially documented · Google Search Central

Microsoft Advertising’s guidance says assistants such as Copilot break pages into smaller pieces that are judged for authority and relevance, then assembled into answers. It recommends clear titles and headings, question-and-answer blocks, lists and tables, and concise, self-contained phrasing.

Officially documented · Microsoft Advertising Blog

For each page: put the direct answer in the first paragraph, state facts with numbers and conditions (who it suits, price range, limits), and say when your product is the wrong choice. The editor’s checklist is in For content editors: writing articles AI answers can quote.

Weeks 6–12: get described correctly by third parties

When people ask “what is the best X”, answers often lean on pages that already list and compare options. Your own site can’t fill that gap alone.

In DATA17’s August 2026 sample of one AI assistant (72 prompts, 297 citations), 80.5% of cited sources were third-party sites. The four most cited sources were all structured category lists.

Observed in practice · observed 2026-08 · DATA17 AI Research Lab · 解数智能 (unpublished report, not publicly available; readers can't check it)

Our own re-test on consumer ChatGPT, in another category, found vendors’ sites made up most citations. The share varies by assistant and category, which is why your baseline records the cited sources (Claim: seeding brand mentions across the web boosts AI visibility).

Use your baseline to find which list, review and comparison sites the answers cite in your category. Then:

• Submit or update your listing on the relevant ones, using your agreed one-line description and accurate prices and features.
• Offer reviewers real material: trial access, data, named customer cases.
• Fix outdated descriptions. A wrong listing can be worse than none.
• Answer real questions in communities as a named company representative.

Google says its AI features can show what is said about products across blogs, videos and forums. It adds that seeking inauthentic mentions is less helpful than it seems, because its ranking systems focus on quality and other systems block spam.

Officially documented · Google Search Central

You can’t write your way onto someone else’s shortlist. You can make it easy for them to describe you accurately.

Weeks 10–13: re-measure and report

Run the same question list, in the same products, with the same number of repeats. Add the search engines’ own reports.

Search Console’s Generative AI performance report shows impressions of your links in AI Overviews and AI Mode, grouped by page, country, date or device.

Officially documented · Google Search Console Help

Bing Webmaster Tools’ AI Performance report counts citations in Copilot and Bing AI answers. Microsoft: “AI Performance does not measure rankings, authority, performance, or importance. It simply shows which content was cited.”

Officially documented · Microsoft Bing

What to report at day 90 is covered in For marketing leads: measuring GEO and reporting it upward. In short: the baseline, what was fixed, what changed in descriptions and citations, and what you plan for months four to six.

What not to do

• Don’t buy mentions or seed fake reviews. Google treats manipulating AI answers as spam, and paid or fake placements are a target in China too (Claim: seeding brand mentions across the web boosts AI visibility, GEO regulation in China).

Google’s definition of spam includes “attempting to manipulate generative AI responses in Google Search.”

Officially documented · Google Search Central

• Don’t mass-produce pages for every question. Scaled pages made to catch queries are what Google’s scaled content policy targets (Scaled content abuse).
• Don’t chase quick technical fixes. llms.txt and extra schema are not documented levers for Google’s AI features (Claim: llms.txt improves AI search visibility, Structured data and AI search).
• Don’t judge the quarter on one screenshot. Answers vary from run to run.
• Don’t cut normal SEO. Google’s AI features use the same index and ranking systems (How Google Search works).

How you’ll know it’s working

R3’s 2026 advertiser white paper advises keeping a GEO pilot narrow, defining a baseline before it starts, and running it long enough to see trends rather than short-term swings. This is consulting advice, not a measured finding.

Hypothesis · R3（胜三） (unpublished report, not publicly available; readers can't check it)

At day 90 the honest signals are:

• A baseline and a second round on the same prompts, each with stated run counts.
• Fewer wrong or outdated descriptions of your brand in answers.
• Your site or listings appearing among cited sources for some problem-type questions.
• Search Console and Bing reports showing AI impressions or citations for the reworked pages.

Category-wide “best X” questions usually move last. Keep the same panel through months four to six.

Measure the same questions the same way every round. A changing panel can show anything you want.

Evidence at a glance

Recommendation	Evidence
Spend the first quarter on baseline and foundations	Hypothesis (industry frameworks)
Run 30+ repeats per prompt	Observed in practice
Confirm indexing and crawler access first	Officially documented
Rework a few core pages for clarity and originality	Officially documented
Get accurately listed on third-party comparison pages	Observed in practice
Avoid inauthentic mentions	Officially documented
Use Search Console and Bing AI reports	Officially documented
Keep pilots narrow with a baseline	Hypothesis