To prove GEO ROI to leadership, show three things in order. First, that each fix changed how often AI answers name you, net of your competitors’ change and beyond its margin of error. Second, that AI assistants sent visits and conversions you can see in your analytics. Third, what the work cost. Present the return as a floor: the part you can measure, with what it misses stated.
GEO here means generative engine optimization: the work of getting your brand named and cited in AI answers. The measurements come from AI search analytics.
Why a rising chart will not convince anyone
Leadership has seen dashboards go up before. Three objections come quickly, and all three are fair:
- “Would it have gone up anyway?” A model update can lift every brand in the same week.
- “Is that a real change?” AI answers vary from run to run, so small moves are often noise.
- “Did it bring in any business?” Being named is not revenue.
A proof of ROI answers each one with evidence. The steps below are in that order.
Step 1: Record the baseline before you start
Before the first fix ships, record 30 days of:
- Visibility and share of voice per engine, for you and your main competitors, each with a margin of error.
- Sessions and conversions from AI assistants in your analytics. How to measure AI search traffic in GA4 shows how to set that up.
Without a baseline, every later number is a claim with nothing to compare it to.
Step 2: Tie each fix to the questions it targets
For every change you ship, such as a rewritten page, a corrected claim or a listing on a site the engines read, write down the buyer questions it is meant to affect. Do this before you look at results. Choosing the questions afterwards lets you pick the ones that happened to rise, and leadership will suspect exactly that.
Step 3: Get a verdict for each fix
This answers the first two objections. For each fix:
- Wait for first retrieval. The fix cannot work until an engine has read the changed page. Start the after-window on the first day an engine cites or retrieves it, not on the publish date.
- Measure the change on the target questions: the after-window against the 30 days before publishing.
- Subtract the field. Take the median change of your tracked competitors on the same questions over the same days. Your net effect is your change minus that median.
- Compare with the margin of error. Above zero by more than the margin is confirmed. Below zero by more than the margin is negative. Inside the margin is inconclusive.
How to tell whether an AI search fix worked covers each step, and whether a change is real or noise covers the margin.
Worked example (illustrative). After a pricing-page rewrite, your visibility on 12 pricing questions rises 8.0 points. The median competitor rose 1.5 points over the same days. Net effect: +6.5 points, with a margin of ±4.0. That clears the margin, so the verdict is confirmed. The numbers are fictional.
Step 4: Count the visits and conversions
This answers the third objection. In your analytics, group visits referred by AI assistants into one channel and read the sessions, landing pages and conversions. In GA4 the Session source dimension holds the referring site.
Know what this misses before you present it:
| Not counted | Why |
|---|---|
| Clicks from Google AI Overviews and AI Mode | Google says they are counted within overall search traffic |
| Visits with no referrer | They land in Direct |
| Buyers who come back later | They arrive by search or by typing your address |
| Buyers who read the answer and never click | No visit, nothing to count |
Every one of these gaps makes the measured figure smaller than the truth, never larger. That is what makes it safe to present as a floor.
Step 5: Work out the return, as a floor
Use leadership’s own numbers for the value of a conversion: your usual rate from demo request to customer, and your usual contract value, from your CRM. Then:
ROI floor = (value of the added conversions from AI-referred visits − cost of the work) ÷ cost of the work
Cost includes the tool, the hours spent and any outside content or agency fees.
Worked example (illustrative). In the baseline quarter, AI assistants sent 310 sessions and 9 demo requests. In the quarter after the fixes, 520 sessions and 16 demo requests: 7 more. Your CRM says one demo request in four becomes a customer, at an average first-year contract of 12,000 in your currency. Added value: 7 × 0.25 × 12,000 = 21,000. The quarter’s work cost 15,000. ROI floor: (21,000 − 15,000) ÷ 15,000 = 40%. The numbers are fictional.
Say plainly what kind of evidence each part is. The verdicts are the causal evidence: they show the fixes raised how often you are named, net of the field. The rise in visits and conversions is consistent with that, but analytics alone cannot prove the fixes caused it.
What to put on the slide
| Line | Example (illustrative) |
|---|---|
| Fixes shipped, by verdict | 6 shipped: 2 confirmed, 1 negative, 2 inconclusive, 1 awaiting first retrieval |
| Best confirmed effect | Pricing questions: +6.5 points net of competitors, margin ±4.0 |
| Visits from AI assistants | 520 sessions this quarter, 310 last quarter |
| Conversions from those visits | 16 demo requests, 9 last quarter |
| Cost | 15,000 |
| ROI, as a floor | 40%, counting only visits analytics can see |
| Not counted | AI Overview clicks, visits with no referrer, buyers who came back later |
How to report AI visibility to leadership has the monthly one-page report this slide sits beside.
What goes wrong, and how to check
| Mistake | How to check |
|---|---|
| Claiming all AI-referred revenue as the result of the work | Subtract the baseline quarter. Some of it would have come anyway |
| Before-and-after with no competitors | Compute the same change for the median competitor. If it matches yours, the engine moved |
| Counting from the publish date | Find the first day an engine cited the changed page. Days before it measure nothing |
| Reporting only the wins | List every fix with its verdict. Negative and inconclusive results make the confirmed ones believable |
| Reading a trend from a handful of conversions | 3 conversions becoming 5 is not a trend. Use a quarter, and show the counts |
| Presenting the floor as the total | Put the “not counted” line on the slide |
| Extrapolating the uncounted part | Do not multiply the floor by a guess. State the gap and leave it unpriced |
An ROI figure leadership can check is worth more than a bigger one it cannot.
How Citino does this
Citino is an AI search analytics tool. It asks ChatGPT, Google AI Overviews and Perplexity your buyers’ questions every day, with Gemini, Google AI Mode and Claude as add-ons. It supplies the first two parts of the proof. It does not calculate ROI, because it has no view of your costs or your CRM.
- Verdicts. When you publish a fix, Citino records the date, waits for an engine to retrieve the page, and measures the prompts the fix targets. The before-window is the 30 days before publishing; the after-window starts at first retrieval. It subtracts the median change of your tracked competitors over the same prompts and days, and compares the result with a margin of error bootstrapped over prompts. Each fix is shown as Awaiting first retrieval, Measuring, Confirmed, Negative or Inconclusive. A page no engine ever retrieved is inconclusive, never negative. The methodology sets this out.
- Visits and conversions. Citino connects to one GA4 property with read-only access and imports aggregate rows only: date, source, landing page, sessions, conversions and revenue. It matches AI-referred visits to the prompts whose answers cited the landing page. This match is a heuristic, labelled as one, and visits it cannot match are shown as unmatched. See Attribution.
Verdicts and GA4 attribution are on every plan; plans differ only in the number of prompts. See pricing.
Everyone tells you what to fix. Citino tells you whether it worked.