1. The basics: what are we measuring?
AI visibility is your brand being mentioned and cited inside the responses that generative AI systems produce. In classic search, success meant “ranking near the top of the results page”; here, success means “being in the response itself.” That difference changes the units of measurement too.
| Question | Classic search | AI response layer |
|---|---|---|
| What does the user see? | Ten links | A single response, 3-5 brands |
| What is measured? | Position, clicks, traffic | Mentions, share of voice, citation share |
| Does traffic follow? | Yes | Most of the time, no |
| Is the result stable? | Relatively | No — the same question can return a different response |
Every method in this guide rests on a single assumption: don’t decide before you measure. Generative models are non-deterministic, so one-off manual checks mislead; look at the sum of many queries and at the tendency over time.
2. Access first: can AI bots reach your site?
The first step of visibility work isn’t content, it’s access. If you block the response bots, no amount of good writing will get you cited in that platform’s response.
The critical distinction is this: training bots and search/response bots are different agents. GPTBot is for model training; OAI-SearchBot is what lets ChatGPT show your site in its search results. Blocking both at once is a common and costly mistake.
- 1Open your robots.txt file and confirm the response bots aren’t blocked.
- 2Check your CDN and WAF rules — Cloudflare bot management can block independently of robots.txt.
- 3Look through your server logs for 403 responses returned to AI bot agents.
- 4Confirm your content is rendered server-side; text produced only by JavaScript can’t be read by some bots.
- 5Accept that content behind a login wall can never be cited.
You can do this step in 30 seconds with our free AI Bot Access Checker; it reports the status of 14 bots one by one.
3. Baseline measurement: know where you stand
The baseline measurement is the reference point for everything that follows. Producing content without it means never knowing which piece of work paid off.
How do you build a question set?
Your questions should represent the buying journey. Asking only discovery questions of the “best X firms” type measures you at the start of that journey and nowhere else.
| Intent | Example question | Share of the set |
|---|---|---|
| Discovery | Which firms are the best for {hizmet} in {sehir}? | ~30% |
| Comparison | X or Y — which one is the better fit? | ~25% |
| Purchase | What are {urun} prices, and where do you buy them? | ~20% |
| Troubleshooting | How is {sorun} solved, and who do you turn to? | ~15% |
| Reputation | Is {marka} trustworthy? | ~10% |
- For a meaningful baseline, 20-40 questions are enough; below that it is statistically noisy, above that unnecessarily costly at this first stage.
- Define 3-5 competitors — not all of them, only the ones you know you are genuinely compared against.
- Always enter your brand’s spelling variants (aliases) — a missing variant makes you look worse than you are.
- Start with at least three platforms; a single platform produces systematic bias.
4. Persona depth: what the average score hides
AI recommends different brands depending on the context of the person asking. When a price-sensitive business owner and a corporate procurement manager ask the same question, they get different lists. Measuring with a single prompt makes that difference invisible.
Your overall score may be 65. But that score may come from an 85 with corporate buyers and a 20 with first-time founders. The average hides the segment you are losing.
Starting with five personas is enough for most brands: two main buyer profiles, one price-focused, one technical evaluator and one skeptic. Widen it as the signal gets clearer.
A cost warning: every persona re-runs your entire question set. 20 questions × 5 personas × 4 platforms = 400 queries. Work out the cost before you widen the scope.
5. Accuracy: it comes before visibility
Being more visible with wrong information costs more than not being visible at all. That is why the knowledge audit comes straight after the measurement.
- 1Put your brand’s verified facts in writing: founding details, products, service regions, certifications, pricing logic. 15-30 items are enough.
- 2Ask targeted questions for every fact and store the responses.
- 3Classify the discrepancies: correct, missing, wrong or outdated.
- 4Fix the source for every discrepancy — the relevant page content, the schema markup and any external references.
- 5After the correction, ask the same questions again and verify the change.
Managing expectations: on platforms that search the web, a correction can show up within days. For information baked into model weights, it takes months. That is why we correct the source content and the structured data together.
6. Content: writing to be cited
When AI assembles a response, it prefers the structured, verifiable sections that answer the question directly. Long introductions, repeated keywords and marketing language are noise in that light.
- Answer the question in the first 40-60 words.
- Use tables, numbered lists and definition blocks; the plain paragraph is the least cited format.
- Add numerical data and dates; present information, not claims.
- Make your headings look like the questions users actually ask.
- Write every section so it still makes sense when it is pulled out of context.
What to avoid
- Producing numerical claims with no source.
- Copying the same content onto different pages.
- Saving the answer for the end of the page.
- Explaining with images alone; information with no text equivalent goes unread.
- Comparisons that run competitors down; models keep their distance from biased content.
Copy a section and cut it away from its context. Read on its own, is it still accurate and clear? If not, AI can’t cite it with confidence either.
7. The loop: measure, fix, measure again
GEO isn’t a one-off project, it’s an operating rhythm. Publishing content is an assumption; asking the target questions again on a regular basis and watching whether your brand gets mentioned turns that assumption into knowledge.
| Rhythm | What happens | Who it suits |
|---|---|---|
| Weekly scan | Score and competitor changes are tracked | Fiercely competitive industries |
| Every two weeks | The same, at a lower cost | Most brands |
| Monthly audit | Factual accuracy is checked again | Everyone |
| Quarterly | The question set and persona set are reviewed | Everyone |
Look at the trend, not the absolute value. What matters isn’t whether your score is 61 or 64, but which direction it moved in over three months.