What is generative engine optimisation?
Generative engine optimisation, usually shortened to GEO, is work intended to improve how a business and its content appear in AI-generated answers. That may involve correcting product information, making evidence easier to find, or improving a page that answers a buyer’s question.
A programme needs a definition of success. Being mentioned, being cited, attracting a qualified visit and gaining a customer are different outcomes. Choose the outcome you can measure, and keep the steps between them visible.
The 2024 GEO research paper presents a framework for studying visibility in generative answers. Its experimental results are evidence about the systems and conditions tested. They do not establish a universal recipe for every current platform or business.
Keep the search fundamentals.
GEO overlaps with SEO because an answer may draw on indexed web pages. Google’s current guidance says its generative search features rely on its core search systems. Pages must meet eligibility requirements, and inclusion is not guaranteed.
Google also says there is no special schema markup required for these features. It does not use llms.txt to improve Google Search visibility. Those statements describe Google Search; do not extend them to every assistant without checking that platform’s guidance.
Our recommendation is to start with accessible, accurate pages that answer a real decision. Treat a content format or publishing tactic as a hypothesis to test, not an established ranking factor for every engine.
1. Choose one audience and one decision.
Write down who is buying, what they need to decide and which market you serve. “Improve AI visibility” is too broad to assign. “Understand how UK agencies compare client-reporting options” gives the team a defined research task.
Collect questions from approved customer research and sales conversations. Label questions you invent as hypotheses. Build a small portfolio spanning discovery, requirements and comparisons. Keep branded questions distinct from questions that do not supply a preferred vendor.
Choose the platforms your investigation needs and record the choice. Keep the portfolio stable through the first review period. The prompt-portfolio guide provides an illustrative six-question set and a record you can reuse.
2. Read the answers before proposing edits.
Capture the question, answer, platform, market, date and source URLs. Check whether the brand appears, what the answer claims and which sources support it. Review the exact source pages instead of assuming every cited domain played the same role.
Separate three kinds of finding: a factual error, a missing explanation and a measurement limitation. A factual error needs verification and correction. A missing explanation needs useful content. A failed collection needs a data check, not a rewrite of the website.
Compare competitors on the same questions. Avoid copying their structure merely because they were cited once. The useful question is which buyer requirement the answer raised and whether your evidence addresses it.
3. Turn one gap into a content brief.
This is an illustrative plan for a fictional reporting-software business. It is not a customer result or a claim that these edits increase citations. The buyer question is: “Can a UK agency give each client a separate reporting workspace?”
Assume the existing page says “reporting for agencies” but never explains client separation. The task is to make the real workflow and limits understandable. Verify the capability before writing; if it does not exist, state that limitation rather than adding a promise.
| Brief field | Example decision |
|---|---|
| Reader | Operations manager choosing reporting software for ten clients |
| Question | Can each client see only its own reporting? |
| Evidence to obtain | Verified access rules, a permitted product demonstration and current plan limits |
| Page change | Explain account setup, permissions and the limits with a concrete workflow |
| Proof to include | A labelled demonstration and a link to the current setup documentation |
| Owner and review | Product verifies the facts; content owner publishes and records the date |
| Measurement | Review the unchanged prompts; inspect claims and sources; record qualified visits separately |
4. Review changes without inventing causation.
Keep a change log alongside the tracked answers. Record publication dates, product changes, altered prompts, platform changes and collection gaps. A useful report makes these visible before interpreting a trend.
For example, suppose a fictional sample moves from six matching responses out of ten to seven out of ten. That is a change from 60% to 70%, or ten percentage points. It is one additional matching response. It does not prove that a page edit caused the change or that sales increased.
Check the underlying answers over repeated observations. If you have enough comparable topics, keep some unchanged as a reference group. Even then, explain the limits of the comparison: platforms and external sources can change during the same period.
Heading provides prompt tracking and answer evidence, with objectives and tasks for the follow-up work. Its visibility metric has a specific denominator. Read that definition before combining it with traffic or conversion results.
A four-week starting plan.
This is a suggested operating plan, not a promised time to results. Adjust the pace to the evidence and people available. The useful output is a record of questions, findings, verified changes and remaining uncertainty.
| Week | Work | Reviewable output |
|---|---|---|
| 1 | Choose a buyer decision, prompts and platforms | A portfolio with sources and a baseline record |
| 2 | Read answers and check their cited pages | A short list of verified gaps, each with evidence |
| 3 | Complete one bounded content or access improvement | A reviewed change and publication date |
| 4 | Revisit the same sample and inspect outcomes | A report separating delivery, observations and hypotheses |
- Do not manufacture reviews, statistics or third-party endorsements.
- Do not create near-duplicate pages for every phrasing of one question.
- Do not treat a crawler visit or citation as proof of a conversion.
- Do not promise a particular position or a guaranteed recommendation.
Start with a question you can investigate.
Choose one buyer decision and save a small set of questions that represent it. Read the answers, follow the sources and assign one improvement whose facts you can verify. Then return to the same evidence at the next review.
The programme succeeds as a learning process before it can be judged as a growth channel. Report what was delivered, what changed in the observations and what is still unknown. Keep commercial outcomes separate until you have the evidence to connect them.
