How to launch a 90-day GEO plan: What do AI models actually cite?

The short answer

Launching a 90-day GEO plan means testing whether AI models cite your content across three measurable phases: a baseline assessment (day 1 to 20), an optimisation cycle (day 20 to 60), and a validation cycle (day 60 to 90). Select 15 to 25 questions your audience is likely to ask ChatGPT, Perplexity or Google AI Overviews, record the sources currently being cited, improve your content structure and authority, then measure again. The goal is not one viral article. It is a repeatable process that shows which changes genuinely improve your chances of being cited before you commit ongoing budget and content resources.

Key takeaways

• Researchers at Princeton and IIT Delhi found that targeted content optimisation can increase a source's visibility in generative AI answers by up to 40% (arXiv, Princeton University / IIT Delhi).
• According to Pew Research Center, around 18% of Google searches produced an AI summary in March 2025. When users saw one, they clicked through to a traditional result in only 8% of visits, compared with 15% when no summary appeared (Pew Research Center via Search Engine Roundtable).
• In the same study, 88% of AI summaries cited three or more sources, yet only 1% of all visits resulted in a click on a link within the summary. Being cited without receiving a click is the reality you need to plan for.
• Work in three-week sprints with a consistent set of 15 to 25 questions. This makes it easier to link changes in citation behaviour to specific improvements rather than random variation.
• Schema.org, launched jointly by Google, Microsoft, Yahoo and Yandex in 2011, remains a core layer for helping both traditional search engines and generative AI systems interpret content (llmoptimisation.fr).

Why do most teams struggle to prove GEO results?

Most marketing teams new to GEO test one article, check ChatGPT a week later, see whether their brand appears, and draw a conclusion. That is not a test. It is guesswork. The issue is not a lack of motivation. It is the lack of a consistent testing routine, a fixed set of questions, regular measurement points and a clear definition of what counts as a citation.

As a result, marketing managers new to GEO run into four recurring problems. First, they have no baseline. If you do not know what AI models cited before you made changes, you cannot attribute later results to your work. Second, content is often written for Google rather than for the way a language model assembles answers from individual passages. That calls for a different writing style. Third, there is no structured follow-up. An article gets published and is never updated, even though GEO requires repeated testing and refinement. Fourth, there simply is not enough time. Writing, measuring and revising content takes weeks, on top of a marketing manager's regular workload.

There is also regulatory uncertainty. From 2 August 2026, Article 50 of the European AI Act introduces transparency obligations requiring providers and users of generative AI systems to tell people when they are interacting with AI or viewing AI-generated content (artificialintelligenceact.eu). This does not change how you optimise for citations, but it does reinforce a key point: AI-driven search is now mainstream and regulated, not a passing experiment.

What goes wrong when you treat GEO as a standalone SEO project?

Many teams approach GEO using the same tools and cadence as traditional SEO: a keyword list, a handful of articles and a quarterly report. That falls short for three reasons.

First, traditional SEO tools measure rankings, not citations. An article can rank third in Google and still never appear in a ChatGPT answer because the model weighs different signals, including structure, how directly the content answers a question and demonstrable authority. Second, the measurement cadence is too slow. AI models are updated regularly, and citation behaviour can shift with each update. A quarterly report misses those changes entirely. Third, teams often fail to connect AI visibility with what is happening on their own site. You may know that Perplexity mentions you, but without connecting that information to Search Console data, you cannot tell whether it is contributing to traffic, enquiries or revenue.

The Pew data makes the stakes clear: users already see an AI summary in 18% of searches, and click-through behaviour within those summaries is extremely low (Pew Research Center). Citations without clicks are becoming the norm. Your brand needs to be mentioned even when those mentions rarely drive direct traffic. If you do not measure that, you are optimising in the dark. That is why a one-off test tells you very little, and why you need an ongoing 90-day GEO plan rather than a single campaign. For a broader view of the approaches that still hold up in 2026, see which AI citation optimisation tactics still work in 2026.

How do you set up a 90-day GEO plan?

A GEO testing programme is a repeatable process of measuring, improving and measuring again, not a one-off content task. A 90-day GEO plan works best when divided into three phases of roughly three weeks each, with review points between them.

Phase 1: Baseline assessment and question research (day 1 to 20)

Start by compiling 15 to 25 realistic questions your audience might ask ChatGPT, Perplexity or Google AI Overviews. Include comparison questions, pricing questions and queries such as "best option for...". Test them manually or with a monitoring tool, then record the answer to each question: Is your brand mentioned? Which sources are cited instead? In what order do they appear? This is your baseline. Without it, you cannot prove in week 10 that your changes made a difference. Connect this phase directly to GEO optimisation as a discipline, so you measure what gets cited from day 1 rather than trying to reconstruct it later. A more detailed guide to this test phase is available in testing GEO in 90 days: how to measure what AI models cite.

Phase 2: Content structure and authority optimisation (day 20 to 60)

In this phase, update content based on what your baseline revealed. The researchers behind the GEO-BENCH framework at Princeton and IIT Delhi found that focused improvements, such as adding citable statistics and clear source references, could increase a source's visibility in generative answers by up to 40% (arXiv). In practical terms, rewrite answers so they can stand alone as citations: make a direct claim, then support it with evidence. Add structured data through Schema.org markup, and make sure every article names its core entity clearly and consistently, whether that is your brand, product or service.

Phase 3: Validation and scaling (day 60 to 90)

Repeat exactly the same question set from phase 1 and compare the results. Which questions now generate a citation where none existed before? Which changes correlate with that shift? This is the point at which you decide which tactics deserve to be rolled out across your wider content cluster, rather than remaining limited to a handful of test articles.

Get started yourself:

• In week 1, create a fixed set of 15 to 25 test questions and repeat them exactly in week 9. Do not change the questions in between.
• Record every source mentioned for each question, not just whether your own brand appears.
• During phase 2, add at least one direct, citable answer paragraph of 40 to 60 words to every article.
• Add Schema.org markup (Article, FAQPage, Organization) to every test article before measuring again.
• Set a threshold in advance: only scale a tactic if it produces a measurable shift for at least one-third of your test questions.

What is the best way to compare GEO tools and approaches?

When you are just getting started, it is easy to get lost among dozens of tools all claiming to measure "AI visibility". Before choosing one, establish a clear evaluation framework. A useful starting point is which comparison actually helps you choose the right SEO tool, which compares the key criteria: coverage across AI models, integration with real site data and language support.

In practice, your decision comes down to three questions. Does the system measure citations across multiple AI models or only one? Does it connect that data to what is actually happening on your site, such as Google Search Console figures, or does it only provide isolated snapshots? And can the platform do more than measure, by improving the content itself based on the data? Most standalone monitoring tools only handle the first task. Launchmind combines measurement and publishing: content is optimised for both Google and AI search engines in one workflow, then refined using real Search Console data rather than a separate dashboard nobody checks every day. See our success stories for practical examples.

How do you stop your 90-day GEO plan from becoming a collection of disconnected articles?

A common mistake is to treat test articles as isolated pieces: one article on topic A, another on topic B, with no connection between them. AI models, however, are more likely to cite sources that build authority across an entire subject area rather than scattered fragments of content. A hub-and-spoke structure, where a pillar page links to in-depth supporting articles that reinforce one another, makes it more likely that a model will recognise your domain as a trustworthy source on the topic.

Imagine a marketing manager at a 40-person business-to-business software company. In week 1 of the 90-day GEO plan, she tests 20 questions around her product category and finds that her brand does not appear in a single ChatGPT response, while two competitors each appear three to four times. By week 6, she has rewritten eight articles into a connected content cluster with citable answer paragraphs and Schema.org markup. She tests the same 20 questions again, and her brand now appears in six of them. It is not a revolution, but it is a measurable, attributable shift she can show to leadership, including the specific changes that contributed to it.

This is where manual work starts to break down. Rewriting eight articles, adding markup and repeating the process every three weeks takes time most marketing teams simply do not have. An AI colleague that publishes daily, builds topic clusters and adapts using Search Console data, such as Alex, can take on that repetitive work without compromising quality. Every article still goes live only after your approval, with a Google preview sent by email.

Frequently asked questions

How much time does it take to set up a 90-day GEO plan?

Creating the question set and completing the baseline assessment usually takes two to three working days spread over the first two weeks. The biggest time commitment comes in phase 2, when you rewrite and restructure content. Depending on the size of the cluster, this can take several weeks if completed manually.

Which tools can I use to test what AI models cite?

You can start by testing manually in ChatGPT, Perplexity and Google AI Overviews using a fixed question list, then track the results in a spreadsheet. To scale across multiple models and languages, you need a platform that monitors citations and connects them to your own content publishing workflow, so measurement and improvement become one process rather than two separate tasks.

Why does my brand show up in Google but not in ChatGPT answers?

AI models weigh different signals than traditional ranking systems. Direct answerability, structured data and repeated, consistent mentions of your brand in context can carry more weight than a strong search ranking alone. Schema.org markup, launched jointly by Google, Microsoft, Yahoo and Yandex in 2011, helps both systems understand the nature of your content, but it does not completely remove the difference in how they weigh signals (llmoptimisation.fr).

Should I be concerned about AI regulation during my GEO programme?

From 2 August 2026, the European AI Act does require transparency around AI-generated content for end users, but it does not classify generative AI such as ChatGPT as a high-risk system (European Parliament). For your GEO strategy, this mainly means being transparent about your own content process. It does not restrict your optimisation methods.

How can I measure whether my GEO work is actually increasing visibility?

Measure three layers at the same time: whether your brand appears in AI answers for your fixed question set, whether your articles perform better in Google Search Console for related queries, and whether that combination leads to measurable traffic. For more detail on tracking brand mentions specifically in ChatGPT, measuring ChatGPT brand mentions is a useful next read.

Conclusion

Launching a 90-day GEO plan is not a marketing gimmick. It is the most reliable way to find out whether your content is genuinely being cited by ChatGPT, Perplexity and Google AI Overviews, rather than guessing from a few isolated screenshots. The three phases, baseline assessment, optimisation and validation, create a repeatable process that lets you adjust every three weeks using real data instead of gut instinct.

The biggest obstacle is rarely knowledge. It is capacity. Rewriting eight to twenty articles, adding markup, repeating question sets and documenting results takes time that most marketing teams cannot consistently free up alongside their day-to-day work. Want to run this process without taking over your calendar? Book a no-obligation call to discuss what a 90-day GEO plan could look like for your brand, or view pricing to see what continuous GEO optimisation costs in practice.

About the company

Launchmind is the AI colleague that writes, checks and publishes SEO content on your own blog every day, in 8 languages, and continuously improves it using real Search Console data. The company supports marketing managers, business owners and chief marketing officers at small and mid-sized businesses and growing companies who know content works but cannot consistently make time for it.