You type your brand name into ChatGPT and it does not appear in the first three answers. You try DeepSeek, then Perplexity, and get the same result. The obvious next step is to buy a monitoring tool, look at the pricing page, and close the tab.
This guide is for the period before the budget is approved. It covers what free methods can actually measure, how to record results so they stay comparable, how much confidence the numbers deserve, and the three points where free checks stop working. By the end you will have a 30-minute weekly routine you can run yourself, plus a clear signal for when to hand the job to a paid platform.
Free checks answer three questions and one they cannot
Free methods are enough to tell you whether a problem exists, not enough to tell you how much it improved.
The first question they answer is whether your brand appears at all for your core prompts. The second is who does appear alongside it. The third is whether AI crawlers are reaching your pages. Those three answers decide whether GEO work is worth starting and which page to fix first.
The question they cannot answer is trend. With a dozen observations, one extra mention this week may come from your changes or from the model answering differently on a Tuesday. That distinction matters: free methods are for diagnosis, monitoring platforms are for verification.
Method one: manual prompt testing with a fixed protocol
The value of manual testing is not accuracy, it is turning "what to ask, where to ask, how to record" into a repeatable protocol.
Manual checks without a fixed protocol produce results that cannot be compared across weeks, and most teams abandon them by week three. Lock down four things first:
- Prompts. Write 5 to 8 fixed questions covering four types: brand ("what is X like"), category ("who does this kind of work"), scenario ("how do I solve this specific problem"), and comparison ("A versus B"). Once set, do not change them for a quarter.
- Platforms. Cover at least ChatGPT, Perplexity, and Gemini for English, plus DeepSeek, ERNIE, and Qwen if you sell in China. Keep the list fixed too — adding a platform mid-stream changes the ruler.
- Fields. Date, platform, prompt, mentioned or not, position, and which page was cited. Six fields is enough.
- Conditions. Same time window each week, same account, same region. Answers vary between runs, so repeat each prompt two or three times and count a mention if it appears in any run.
The whole routine takes about 30 minutes a week and produces a ledger that shows direction. Its ceiling is obvious: small sample, few platforms, no competitor view. As a first step it still beats doing nothing.
Method two: use Search Console impressions as a proxy
Search Console does not count exposure inside AI answers, but it shows which pages already have the corpus foundation to be cited.
AI answers are assembled from pages that can be crawled and chunked. A page that cannot earn impressions in classic search is either poorly indexed or matching no real query, which makes it an unlikely source for an AI answer. So Search Console here is not an AI visibility measurement — it is a filter for pages that qualify to be cited.
Three signals worth watching. First, question-style long-tail queries start earning impressions, which means the page is drifting toward conversational search. Second, impressions rise while click-through stays low, which is common when a page is seen in an answer layer without being clicked. Third, a page with zero impressions over months is not an optimization target yet — leave it alone.
One caveat: Google's documentation states that clicks from AI features are counted within overall Search performance rather than reported as a separate AI exposure metric, so treat this as a proxy only. Google's documentation on AI features in Search is the reference to read before drawing conclusions.
Method three: read the logs to see whether AI crawlers arrive
A GPTBot, ClaudeBot, or PerplexityBot hit in your logs proves the door is open, not that anything was cited.
The point of this step is to rule out basic mistakes. Plenty of sites finish a content overhaul and see nothing change because robots.txt blocks AI crawlers outright, or a CDN rule rejects non-browser user agents. Group hits by user agent, then check whether the pages being fetched are the ones you want cited. Ten minutes covers it.
Keep two things separate: fetching means "they came", citation means "they used it". Logs only prove the first. If crawling is healthy, content is clean, and prompt tests have not moved in three weeks, the problem is usually that the page never states a conclusion clear enough to lift out.
For the crawler access rules themselves, RFC 9309 defines how robots.txt works, and vendor documentation such as OpenAI's bot reference explains each agent's purpose.
Where free stops working: three breakpoints
Free methods fail in three concentrated places: sample size, cross-platform consistency, and competitor comparison.
Sample size. Five prompts run three times is fifteen observations. A two-mention swing before and after a change is indistinguishable from noise. Any real trend claim needs hundreds of observations, which manual testing cannot supply.
Cross-platform consistency. Chinese and Western AI search differ in prompt habits, corpus sources, and ranking logic. The same question produces structurally different answers on DeepSeek and ChatGPT, and running both by hand quickly costs more in hours than a subscription costs in dollars.
Competitor view. Manual testing tells you whether you were mentioned. It does not tell you how many of the five brands named in that answer were your competitors — and that ratio is usually what decides whether the project gets funded.
What each method can and cannot measure
Comparing by dimension rather than by price makes the decision easier.
| Dimension | Manual prompts | Search Console | Log analysis | Monitoring platform |
|---|---|---|---|---|
| Brand mentioned at all | Yes | Indirect | No | Yes |
| Position in the answer | Partial | No | No | Yes |
| Cross-platform coverage | Costly | No | No | Yes |
| Competitor share | No | No | No | Yes |
| Trend and confidence | Too few samples | Indirect | No | Yes |
| Cited source pages | By hand | No | Partial | Yes |
| Automation | Weekly | Daily | Real time | Daily or better |
The table makes one thing clear: free methods cover diagnosis, paid platforms cover verification. They are not substitutes. Diagnose free, verify paid, and you spend the least. If you want a baseline before building the ledger, run AiVisible's free scan, which checks visibility across DeepSeek, ERNIE, Qwen, ChatGPT, and Perplexity and gives you the first row of the table above.
Three signals that it is time to pay
If any of these appears, manual testing has become a waste of time.
First, you need to report a trend to a manager or a client. Trends need confidence intervals, and a hand-kept ledger invites exactly the methodology question you cannot answer. Second, you need competitor comparison. Competitor share within the same prompt set is nearly impossible to capture by hand, and it is often the number that unlocks budget. Third, you are optimizing for three or more platforms at once. The more platforms, the faster manual cost grows and the better a platform's relative price looks.
Conversely, if you only need to confirm whether your brand shows up in AI answers at all, or whether last month's content rewrite moved in the right direction, free methods are sufficient. Those judgments need direction, not precision.
A 30-minute weekly routine
Making the check routine matters more than the tool you choose.
- Run the prompt set at a fixed time on Monday. Eight prompts, three platforms, same order, no ad-hoc additions.
- Fill the ledger. Six fields, mention counts if it appears in any run, position recorded as a rough band (top three, three to five, absent).
- Compare with last week. Look at two things only: prompts newly mentioned and prompts dropped. Ignore the middle.
- Fix one page. Prefer pages that have been cited but bury their conclusion — that change has the shortest path to visible effect.
- Summarize monthly. Merge four weeks into one view and read direction, not single-week numbers. For the scoring logic behind it, see how AI visibility scores are calculated.
This routine is the minimum viable version of a real audit, not a replacement for one. When the output has to be shown to someone else, extend it using the full AI visibility audit process. If you would rather start on the content side, the guide to ChatGPT citation logic covers where a conclusion sentence needs to sit.
FAQ
Are free AI visibility checks reliable?
Reliable for direction, not for magnitude. Free methods confirm whether a brand is mentioned, which competitors appear, and whether crawlers visit. But with only a dozen observations they cannot support a trend claim. Use free checks for diagnosis and paid monitoring for verification.
How many prompts does manual testing need?
Five to eight is enough to start, covering brand, category, scenario, and comparison questions. Consistency matters more than count — changing prompts mid-quarter destroys comparability. When you need trend data, add repetitions rather than prompts.
Can Search Console show traffic from AI search?
Not as a separate field. Google states that clicks from AI features are counted within overall Search performance, so impressions and click-through can only serve as a proxy. It is better used to find pages with a corpus foundation than to produce a visibility score.
Does a GPTBot hit in my logs mean I will be cited?
No. Fetching proves the crawler arrived; citation also depends on whether the page can be chunked, whether conclusions are stated clearly, and whether third-party sources mention you. Healthy crawling with flat prompt results points to content structure, not technical access.
When should I move from free checks to a paid platform?
Three signals: you must report trends externally, you need competitor comparison, or you are optimizing three or more platforms. Any one of them makes manual work more expensive than a subscription. For a one-off baseline or a direction check, free methods remain sufficient.