How do I prove to clients that AI search optimization is working?
Quick Answer: Measure against a dated baseline of fixed tracked questions, then report how the brand's visibility, share of voice and position moved, plus answer tone, cited pages and competitors, beside Google's Search Console generative AI report [2]. A move repeated across consecutive checks and several questions is a trend; one move on one question is noise. This is for agencies billing clients, not brands tracking themselves.
Key takeaways
- Put the baseline date at the top of every ai search optimization report.
- Before you send a Search Console export, relabel zeros that showed as
~or-on screen as "not available". - Agree the tracked question list with the client in writing before the first check runs.
What counts as proof that AI search optimization is working?
Proof is a change you can date. You ask the same tracked questions of the same engines and compare the answers against a baseline taken before the work began. The client report reads six things from those answers: how often the brand appears, its share of mentions against named rivals, where it is named, the tone of the answer, the pages cited, and which competitors appear.
One good ChatGPT screenshot proves nothing. Welt Analytics' about page explains why: "AI engines don't answer the same question the same way twice." [1]
How do you set a baseline before you start any ai search optimization work?
To set a baseline, choose the client's tracked questions once and freeze them. Then run a dated first check on every engine before any new page or article ships.
Welt Analytics' pricing page defines a prompt as one tracked question and states: "The plan limit is the number of questions, not the number of engines." [3] Welt Analytics scans the customer's site and sets the first prompts, and its 7-day free trial needs no card, with 10 prompts [3]. The trial gives a first baseline on up to 10 questions.
The most common mistake is taking the baseline after the work has started. Run and date the first check before any content or page change, and keep the same questions afterwards.
- Open the client's tracked questions in your tracking tool. Confirm the list matches the one the client approved.
- Confirm the baseline date and that all six engines are listed.
- Copy the first results into your report template, one row per engine.
Expected result: one dated baseline row per engine, with visibility, share of voice, position, sentiment, sources and competitors filled in.
Which numbers show a change, and how often do they refresh?
Each engine refreshes on its own schedule, so read every change only against that engine's earlier checks. Welt Analytics checks Google AI Mode, Google AI Overviews, Gemini and ChatGPT daily, and Claude and Perplexity weekly [1].
| Engine or report | Check cadence | How to read a move |
|---|---|---|
| Google AI Mode, Google AI Overviews, Gemini, ChatGPT | Daily | Chart a weekly average so daily swings do not read as wins or losses. Call a trend only after several consecutive checks. |
| Claude, Perplexity | Weekly | Compare week to week only and label the chart as weekly. Keep these points out of any daily line. |
| Google AI features in Search Console | Google's own date grouping (next section) | Treat it as Google's own measurement. Keep it beside the tracked-question view as a separate line. |
A common trap is one chart that plots all six engines on a daily axis. Claude and Perplexity then show flat steps between their weekly points, and a client reads each step as a sudden jump. Label the cadence on each engine's line and compare each engine only against its own earlier checks. To see how the share figure is built against named rivals, read how to measure share of voice in AI answers versus competitors.
How do you show proof on the Google side with Search Console?
Use the generative AI performance report: Google's help page says it "shows data about how your site performs in generative AI features on Google Search." [2] The report groups dates by days, weeks or months in Pacific Time. It also has an export button for the chart and table data, and any value shown as ~ or - on screen becomes a zero in the download.
A page counts only if Google can show it at all. Google's guide says a page must be indexed and eligible to appear in Google Search with a snippet [4]. The same guide adds: "a site must be included in Search generative AI features in Search Console to be eligible for display in generative AI features on Google Search." Google's URL Inspection help says you type the full URL into the inspection bar at the top of any Search Console screen, for the property you have open; URL is on Google means the page is indexed, and URL is not on Google means it won't appear [5]. This check confirms indexing only.
- Pick one client page you expect to see in Google's AI features.
- Inspect it from the inspection bar while the client's property is open.
- Read the status line at the top of the result.
Expected result: the page shows URL is on Google. If it shows URL is not on Google, fix the indexing before you report on that page.
A related trap is giving schema, an llms.txt file or bought mentions the credit for a gain. Google's guide states: "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add." [4] The same guide says you don't need new machine-readable files or AI text files to appear in Google Search. It also says that chasing inauthentic mentions across the web helps less than it might seem. Tell clients that Google says none of these is needed for Google Search, and never bill them as the cause of a gain there.
How do you write the report so a client believes it?
Open with the comparison from baseline to now. Follow it with the sources the engines cite, and finish with the competitors. The layout and the monthly rhythm are covered in how to report AI search visibility to clients.
The trap that hurts most starts in the proposal, when it implies an outcome. Welt Analytics' about page puts it plainly: "Nobody controls what an AI engine says." [1] Promise the client a measured report and a method.
Crediting a gain to one article is the other easy mistake. A daily engine can move for reasons that have nothing to do with the page you shipped. On Welt Analytics' Growth and Pro plans, articles are written and published on a blog on the customer's subdomain. List those articles with their publish dates next to the metric changes.
Where does Welt Analytics fit into the proof pack?
Welt Analytics is an AI visibility tracking and managed article publishing service. It covers the tracked-question side of the pack: six engines on the split cadence above, and the report metrics plus competitors for each engine. Plans as of 2026-11-13 (listed prices exclude VAT):
- Starter: 89 EUR per month. 50 prompts and 1 brand across all 6 engines, with unlimited seats.
- Growth: 229 EUR per month. 150 prompts and 2 brands, the Data API and 1 managed article a day.
- Pro: 499 EUR per month. 400 prompts and 3 brands, the Data API and 3 managed articles a day.
If you are comparing options across a client roster, see the best AI visibility tool for agencies managing several client brands.
To build a first baseline for a client before the retainer's first article goes live, start with Welt Analytics and freeze the question list on day one.
Questions
What if a client's numbers do not move after a month?
Report the flat month as a result. Then look at which sources the engines cite for the frozen questions and which competitors keep appearing, and aim next month's work at those sources. Note what you changed and when, so the next comparison has a reason attached.
Can I add new tracked questions halfway through a retainer?
Yes, as a second set with its own start date. Keep the original set running unchanged so the first baseline stays comparable, and report the two sets on separate lines.
Can I show a client proof without a paid tool?
Search Console covers Google's side. For ChatGPT, Claude, Gemini and Perplexity, you can ask the frozen questions by hand on a fixed day and log the answers in a sheet. That works for a handful of questions and gets slow past that. A hand log also captures one answer per question, while an automated check keeps collecting answers over time.
Who owns the articles if the client ends the plan?
Welt Analytics says every published article belongs to the customer. The customer can export the articles while the plan is active and for 90 days after it ends [6].