
Can free tools give you a useful AI-search baseline?
Build a modest, repeatable starting point using native reports, manual observations and a simple evidence log before buying more software.
Written by Rob KnottYes, free tools and a disciplined manual process can give you a useful starting point. They can help you identify inaccurate answers, gaps worth investigating and questions your team needs to answer. They cannot provide complete visibility into every AI conversation.
The trade-off is time and coverage. This guide gives you a deliberately small baseline you can run before deciding whether paid tooling or specialist support is worthwhile. It is an exploratory method, not a statistically representative survey of all buyers.
Choose one question your business needs answered
Start narrowly. For example: “When a UK operations manager compares maintenance providers, does our business appear, and are our service areas described accurately?” Define the audience, market, offer and decision stage.
Avoid beginning with every product, every platform and every country. A small study with a clear purpose is easier to repeat and interpret. Write down the exclusions so nobody later mistakes a UK service comparison for a global brand assessment.
Decide what would change your next action. An incorrect service area might lead to a factual content review. Repeated references to outdated third-party descriptions might justify investigating those sources. Simply obtaining a larger number is not a useful action by itself.
Keep three evidence streams separate
Use native webmaster reporting where available, observed answers from a documented sample and your own visit or enquiry information. These streams answer different questions. Do not combine them into a percentage with no coherent denominator.
Bing's AI reporting provides information about citations within the Microsoft and partner experiences it covers. Google's documentation explains that traffic from its AI features is included within Search Console's overall Web performance reporting. Neither should be relabelled as a universal report of all AI search.
Analytics can identify some referral visits where source information is available. It will miss or misclassify some journeys, and a person can hear about you in an assistant before visiting through another route. An optional enquiry question such as “How did you first hear about us?” provides useful context without resolving every attribution gap.
| Evidence | Use it for | Keep this limitation visible |
|---|---|---|
| Available webmaster reports | Inspect cited pages or search performance within the stated coverage | Coverage differs by product and can change |
| Documented manual answer sample | Review brand presence, recommendations and factual accuracy | A small sample is sensitive to conditions and variation |
| Analytics referral data | Review identifiable visits and behaviour | Not all AI-influenced visits retain a useful referrer |
| Enquiry context | Learn what customers say influenced them | Self-report is incomplete and not causal proof |
Build a small prompt set from real questions
Gather questions from sales calls, support tickets, enquiry forms and your existing research. Remove personal information. Group the questions by discovery, comparison and validation. Keep branded accuracy checks separate from questions that do not name you.
For a fictional commercial-cleaning company, a discovery question might ask how to find an office cleaning provider in a specific service area. A comparison question could ask what to evaluate before choosing a provider. A branded check could ask whether the company offers weekend work. These are examples of question types, not a universal prompt list to copy unchanged.
Keep the wording natural and the scope relevant. Do not insert flattering attributes simply because they make your brand more likely to appear. Save a fixed core set for later comparison, and maintain a separate list for exploratory questions you add as you learn.
Use the measurement framework for a deeper explanation of sampling and scoring.
Run a manageable pilot and record the conditions
One possible pilot is 12 questions on two selected platforms, with three observations of each combination on separate occasions. That produces 72 observations. It is a workload example, not a minimum sample size or a claim of statistical confidence.
Use fresh conversations to reduce carryover and record the visible platform, date, account/session conditions, language, location settings where controllable and whether web search is active. Personalisation and system changes may still affect results. Follow platform terms; do not assume that a free interface authorises automated collection.
Allow time to inspect the actual answers, not just tick a mention box. At an illustrative three minutes per observation, 72 observations would take 216 minutes before setup and analysis. Your actual time will vary. “Free software” does not mean free research.
| Field | Example or instruction |
|---|---|
| Question ID and category | D04 / unbranded discovery |
| Exact prompt and context | Save the wording and any relevant settings |
| Platform and observation date | Record the visible product and time |
| Response status | Usable answer, failed run or unavailable |
| Brand presence and recommendation | Record separately, with definitions |
| Cited URLs | Save the actual visible sources where permitted |
| Accuracy check | Claim, verified fact and evidence |
| Interpretation and next action | Separate observation from hypothesis |
Worked example: interpreting a small result responsibly
Suppose the fictional cleaning company appears in 9 of 24 usable discovery observations. That is 37.5% of this particular set of observations. It is not 37.5% of the market, all prompts or all potential buyers. Describe the denominator every time you report the result.
You also find three answers that describe a service the business no longer provides. Review the exact wording and cited sources. The company website might be unclear, an external listing might be outdated, or the answer might not provide enough evidence to establish the source of the claim.
Your action could be to correct verified outdated descriptions and make current service boundaries easier to find. Record the changes, then repeat the same core sample under comparable conditions. If the next set improves, report the change and the limitations. Do not claim that a single edit caused all of it.
Turn the baseline into a short action list
Group findings into factual corrections, technical checks, content gaps and questions requiring further evidence. For each action, record an owner, a reason, the intended change and how you will verify implementation.
Review business relevance before expanding the sample. An obscure prompt that produces an amusing answer may be less useful than a repeated error about your main service. Prioritise the issues that could affect a customer's understanding or decision.
Schedule a proportionate recheck after changes have had a reasonable opportunity to become discoverable. Keep your core questions stable while reporting any unavoidable changes in platform behaviour. Baseline comparisons are stronger when you can explain what stayed the same and what did not.
The AI visibility improvement guide helps connect observations to defensible changes.
Know when to buy a tool or bring in help
Paid tools can become useful when collection consumes too much staff time, several markets need consistent coverage, evidence handling becomes unwieldy or repeat reporting needs dependable operations. Evaluate whether a product actually covers your platforms, exports the evidence you need and documents its collection method.
A tool does not decide which business questions matter or whether a finding justifies an action. Specialist help is valuable when those decisions are the bottleneck. You may only need a baseline design, a setup workshop or an independent interpretation of the evidence.
Start with the smallest credible workflow, cost the human effort and review what it taught you. Then invest in the part that genuinely limits your team's progress.
Key takeaways
- A small manual baseline can support decisions if its scope is explicit.
- Keep native reports, observed answers and customer outcomes distinct.
- Record conditions, missing data and definitions so you can repeat the work.
- Buy tooling for a demonstrated operational need rather than a headline score.

