How-tos · November 19, 2026 · 11 min read
A 90-Day GEO Strategy: The Plan and the Checklist
A three-phase GEO plan for a quarter: two weeks of baseline sampling, six weeks of work on your pages and cited sources, a fair comparison, and a checklist.
A 90-day GEO strategy has three phases. In days 1 to 14 you measure: fix a prompt set, pick your competitors, sample the assistants every day, and read where you are missing and which sources they cite. In days 15 to 60 you work: fix your own pages, earn a place in the cited sources, make your brand facts consistent, and clean up crawling and indexing. In days 61 to 90 you measure again with the same prompt set, keep what moved and drop what did not.
Generative engine optimization (GEO) is the work of getting named and cited in AI answers. This post turns it into a quarter’s plan with dates, a checklist you can copy, and advice on what to tell your boss. For the tactics behind each step, see our playbook for improving AI visibility. For how GEO relates to the search work you already do, see GEO vs SEO.
Why 90 days
A quarter is long enough to see results from retrieval and short enough to keep focus. When an assistant searches the web before it answers, a new or improved page can be cited within weeks of being indexed. So changes made inside 90 days can land inside 90 days.
What a quarter will not reliably show is a change in what the models learned in training. That waits for new model versions, which ship on the providers’ schedules, not yours. Plan for the first kind of result inside the quarter, and treat the second as a longer bet.
Days 1 to 14: baseline
The goal of the first two weeks is a number you can trust and compare against at day 90. Do no site changes yet.
Write the prompt set. 15 to 25 questions in the words buyers use. Most should leave your brand name out, because those measure recommendation, not recognition. Cover the main buying moments: “best X for Y”, “alternatives to Z”, “X vs Y”, “how do I solve [problem]”. Our guide to checking whether ChatGPT recommends your brand covers prompt writing in detail.
List the competitors. The 3 to 8 brands buyers compare you with. Write down every alias for each, including yours. Decide what counts as a mention before you start.
Pick the engines. Track the assistants your buyers use, each one separately. At minimum ChatGPT, Perplexity and Google AI Overviews.
Sample every day for two weeks. One run a day for 14 days gives 14 answers per prompt per engine. That clears the floor of 10 answers we treat as the minimum for reading a rate.
Read gaps and citations. At day 14, list the prompts where a rival is named and you are not. Then group the sources cited in those answers: review sites, comparison articles, communities, directories, competitor pages, your site. These two lists are the plan for the next six weeks. Our guide to AI brand monitoring covers how to read them per prompt and per engine.
Days 15 to 60: the work
Work from the two lists, most frequent first. Log the date of every change.
Fix your own pages. For each gap prompt, ask: does a page on our site answer this question directly? If not, write it. If yes, rewrite it so the first paragraph is a complete, quotable answer. Put prices, limits and audience in text. Publish honest comparison pages for the rivals buyers ask about.
Earn presence in the cited sources. Complete your listings on the review sites that keep appearing. Ask real customers for reviews. Contact the authors of the cited comparison articles with corrections or something useful to test. Take part honestly in the communities buyers use, with your affiliation stated.
Make your entity consistent. Write one sentence that says what you are and who it is for. Use it on your homepage, your about page, your social profiles, your review listings and your directory entries. Remove old prices, old plan names and old positioning wherever they still appear. Assistants blend many sources, so conflicting facts produce a blurred answer.
Clean up technical hygiene. Check robots.txt for the AI crawlers you want to allow. Make sure key pages render their main text in HTML. Submit sitemaps to Google Search Console and Bing Webmaster Tools, and confirm your key pages are indexed in both.
Keep the measurement running the whole time.
Days 61 to 90: measure and iterate
Compare against the baseline with the same prompt set. Same prompts, same engines, same schedule, same detection rules. If any of these changed, the comparison is not fair. Compare per prompt and per engine, not just the overall number.
Match moves to changes. For each prompt that moved, check your change log. Did a new page start to be cited? Did a review site you fixed start to appear? A move with a matching cause is probably yours. A move across many prompts on one date, with no matching change, is probably a model update.
Keep what moved, drop what did not. Do more of the kind of work that moved prompts. Stop the work that moved nothing after six weeks. Then plan the next quarter from the new gap list.
The 90-day GEO checklist
| # | Task | Why | How to verify |
|---|---|---|---|
| 1 | Write 15 to 25 buyer prompts | You can only measure what you ask | Sales or support confirms the wording; most prompts omit your brand |
| 2 | List 3 to 8 competitors with aliases | Fair share of voice needs every name | Aliases tested against real answers |
| 3 | Write the mention rule | Same detection rules for every brand | Rule written down before day 1 |
| 4 | Pick engines by where buyers are | Engines disagree with each other | Engine list agreed and fixed |
| 5 | Sample daily for 14 days | One day is noise | At least 10 answers per prompt per engine |
| 6 | Save the baseline with its date | You compare against it at day 90 | Baseline report stored |
| 7 | List gap prompts | Gaps are the targets | Each gap names the rival who appears |
| 8 | Group cited sources by type | Citations show where to work | Top cited domains counted |
| 9 | Start a change log | Attribution needs dates | Every change has a date and a URL |
| 10 | Rewrite key pages answer-first | Assistants quote direct answers | First paragraph answers the title question |
| 11 | Publish honest comparison pages | “X vs Y” questions need a page to cite | Pages live and indexed |
| 12 | Add FAQs with real buyer questions | Each question is a quotable unit | FAQ schema matches the page text |
| 13 | Complete cited review listings | Review sites shape answers | Listing complete, facts current |
| 14 | Contact authors of cited comparisons | Being in the article puts you in the answer | Outreach logged with replies |
| 15 | Write one brand sentence and use it everywhere | Consistent facts give a clear answer | Same sentence on site, profiles and listings |
| 16 | Remove old prices and plan names | Wrong facts spread into answers | Direct-question prompts describe you correctly |
| 17 | Review robots.txt for AI crawlers | Blocked pages cannot be retrieved | Allowed crawlers listed on purpose |
| 18 | Confirm indexing in Google and Bing | Assistants retrieve from search indexes | Key pages indexed in both |
| 19 | Render main text in HTML | Some crawlers do not run scripts | Page source shows the main text |
| 20 | Compare at day 90 with the same setup | Only a fair comparison means anything | Same prompts, engines, schedule and rules |
An llms.txt file is cheap to publish, and you can add it if you like. It does not replace any item above.
What to report to your boss, and when
Set expectations on day 1. Say plainly that the first two weeks produce a baseline, not results. Say that changes which depend on retrieval can show within weeks of a page being indexed and cited, and that changes which depend on model training take longer and arrive when new models ship. Do not promise a number. Nobody can predict how far one change moves an assistant’s answers.
A simple reporting rhythm:
- Day 14: the baseline. Visibility rate and share of voice per engine, the top gap prompts with the rival who wins each, and the top cited sources. This is the “where we stand” report.
- Days 30 and 60: the work. What you changed, with dates. Which changes are already cited. Early movement on specific prompts, labelled as early.
- Day 90: the comparison. Baseline against now, same prompt set, per engine. What moved, what did not, and what you will do next quarter.
Report the sample size next to every rate. Illustrative example: “62 percent (from 14 answers)” invites better questions than “62 percent” alone.
Common failure modes
Changing the prompt set mid-test. A new prompt set measures a different thing. If you add or rewrite prompts at day 40, the day 90 number no longer compares with the baseline. Collect new prompt ideas in a list and add them next quarter.
Reading one day. Assistants give different answers each time. A single day can show a jump or a drop that disappears the next day. Read weekly averages, and never react to one run.
Chasing one engine. The engines retrieve from different sources and disagree with each other. Work from the gaps across every engine your buyers use, and accept that one change may move one engine and not another.
No change log. Without dates, you cannot tell your work from a model update.
How CitePrism runs this plan
CitePrism is a GEO tool built for this loop. It runs your prompt set on a schedule against ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, Google AI Mode and Microsoft Copilot, and scores you and your competitors by the same rules. It stores every raw answer, greys out rates below 10 answers, and counts named and cited separately.
For the baseline phase, the Gaps page lists prompts where a rival is named and you are not, and citation share shows which sources shape the answers. During the work phase, Opportunities gives a weekly action list written from the last 30 days of answers, with each item tied to the prompts it helps. At day 90, the trend lines show per-engine rank and visibility against your baseline.
The Free plan covers one brand, 25 prompts, 5 competitors and 1 run a day on your own keys. The model providers charge you directly: about $0.046 per prompt across the four API engines on their default models (our measurement, August 2026). Twenty prompts once a day comes to roughly $28 a month. The engines that have no API need their own key on Free, priced per request by its provider. Pro is $29 a month, runs on CitePrism’s keys with a daily included usage amount, and allows 4 runs a day and 25 competitors per brand.
FAQ
What is a GEO strategy?
A GEO strategy is a plan for getting your brand named and cited in AI answers. It has three parts: measure where you appear across a fixed set of buyer questions, work on your own pages and on the sources the assistants cite, then measure again the same way to see what moved.
How long does GEO take to show results?
Changes that depend on retrieval can show within weeks, once a new or improved page is indexed and starts to be cited. Changes that depend on what a model learned in training take longer, because they wait for new model versions. A quarter is enough to see the first kind and start the second.
Can I run a GEO strategy without a tool?
Yes. A spreadsheet, a fixed prompt set and daily discipline will do. The hard part is the repetition: 14 days of daily sampling across several engines is a lot of manual logging. A tool runs the schedule and keeps the comparison fair, but the plan is the same.
What should a GEO checklist include?
It should cover four areas: measurement (prompts, competitors, engines, baseline), your own pages (answer-first content, comparisons, FAQs), presence in cited sources (review sites, comparison articles, communities) and technical hygiene (crawler access, indexing in Google and Bing, HTML rendering). Every item needs a way to verify it is done.