Short answer, as of August 2026: build the case on four numbers — the share of your category's questions where you are currently absent from AI answers, the value of a customer, a deliberately conservative influence rate, and the fully-loaded cost of doing the work. Referral clicks belong in the model, but as a floor, not as the whole case. AI answers are designed to reduce the click, which means the click undercounts the effect by construction.
This is the version of the argument that survives contact with a finance team, including the parts where the honest answer is "we don't know yet".
Why the naive calculation fails
The instinct is to open analytics, filter to referrals from the AI assistants, count the sessions, and divide the cost by them. Do that in month two and the number will be laughable — a few dozen sessions against a four-figure investment.
Three things are wrong with it.
AI answers are engineered to satisfy without a click. When an assistant names your business and summarises what you do, the user may act on that — by searching your brand, calling you, or walking in — without ever touching a referral link. The visit that follows is attributed to direct or branded search, and the assistant that caused it is invisible.
Referral attribution is incomplete even when a click happens. Referrer headers get stripped, in-app browsers behave inconsistently, and some assistants surface a source without a live outbound link at all. The measured number is a subset of a subset.
The effect is not linear in time. Citations accrue and then compound: a page cited once becomes a source the engine has already validated. A two-month window measures the ramp, not the plateau.
None of that makes GEO unmeasurable. It makes referral clicks alone the wrong denominator. The plumbing for measuring what can be measured is covered in AI search traffic attribution and analytics; this piece is about the decision on top of it.
The four numbers
1. Absence rate — the size of the gap. Take the questions a genuine prospect asks an assistant before choosing a supplier in your category. Ask each one, across the engines your market actually uses, and record whether you appear. If you are named in 4 of 60 prompt-checks, your absence rate is 93%. This number is free to establish and it is the entire premise of the business case: you cannot be recommended in conversations you are not part of.
2. Customer value. Average order value or first-year contract value, times expected repeat, less delivery cost. Use the number your finance team already uses. Do not invent a new one for this exercise — half the credibility of the model is that it borrows the company's existing arithmetic.
3. Influence rate — deliberately conservative. How many of the people who see you named in an answer become a customer, by any route? Nobody has a defensible industry figure here, and anyone quoting one precisely is guessing. So do not guess: pick a low number, label it an assumption in bold, and show the model at that number and at half of it. A case that survives its own pessimistic column is a case worth funding.
4. Fully-loaded cost. Software, plus the hours. If the tool costs $249 a month and someone spends four hours a week on the work, the hours are the larger line. Leaving them out is the single fastest way to lose the room when someone notices.
A worked example
Assume a regional services business: 60 prompt-checks a month across the six major engines, currently named in 4. Average customer value $1,400. The team invests four hours a week plus a $249/mo platform.
| Line | Month 0 | Month 6 (modelled) |
|---|---|---|
| Prompt-checks where we are named | 4 / 60 | 22 / 60 |
| Estimated monthly answer impressions | ~120 | ~660 |
| Influence rate (stated assumption) | 1.0% | 1.0% |
| Modelled customers influenced / mo | ~1 | ~7 |
| Value at $1,400 each | ~$1,400 | ~$9,240 |
| Platform cost | $249 | $249 |
| Labour, 16 hrs/mo at $75 | $1,200 | $1,200 |
| Net | −$49 | +$7,791 |
Every number in that table is a stated assumption except the first row, which is measured. That is the point of the format: it makes the argument auditable. A sceptical CFO does not have to accept your influence rate — they can halve it, watch the month-six net fall to roughly $3,170, and still see a defensible investment. If halving the assumption destroys the case, you have learned something important before spending the money rather than after.
The "answer impressions" line is itself an estimate — it scales named prompt-checks by how often those questions get asked in your market. Say so explicitly. A model with three honest estimates beats a model with three confident-looking fabrications, because the first one can be revised as data arrives and the second one collapses the moment anyone checks.
The comparison that actually gets asked
The question in the room is rarely "is GEO worth it in the abstract". It is "is this worth more than the next dollar of paid search, or another writer, or the thing we were going to do instead".
That comparison has a useful asymmetry: the highest-value GEO work is not additive spend, it is redirected spend. Fixing crawler access costs an afternoon. Restructuring your ten best existing pages to answer questions directly is the same content budget you already have, spent differently. Making your entity data consistent across directories is tedium, not capital.
Only after those does GEO start needing net-new investment — and by then you have a measured baseline to justify it with. That sequencing is what makes the case winnable: the first phase is close to free, which means the first data point costs almost nothing. We laid the budget split out in more detail in GEO versus traditional SEO: how to split the budget.
Worth being explicit about the overlap, too: most of what earns AI citations — clear structure, specific verifiable claims, consistent entity data, third-party corroboration — also helps classical search. You are rarely choosing between the two so much as choosing what the same work optimises for.
What to report, and how often
Three charts, monthly, on one page.
Citation rate over time, per engine rather than averaged. Averaging hides the interesting event, which is one engine flipping while the others hold.
Share of voice against named competitors. "We are named in 22 of 60 prompt-checks, the market leader in 41" is a sentence an executive can act on. A raw score out of 100 with no comparator is not.
Measured referral and branded-search movement. This is the floor of the case, presented as the floor. Show AI referral sessions alongside branded-search volume, because a rising branded-search line during a citation push is the fingerprint of the unattributable half of the effect.
Then one paragraph of what changed and what you are doing next. Resist the temptation to add a fourth chart; the discipline of three is what makes it get read. If you are producing this for clients rather than internally, white-label SEO reporting for agencies covers the packaging.
The cost side people forget
Two lines routinely go missing from GEO budgets, and both show up later as unpleasant surprises.
Tool sprawl. Citation tracking, content scoring, rank tracking, directory audits and a reporting layer bought separately run to several hundred dollars a month before anyone has written a word — and the integration cost is a human doing exports. We put numbers on the assembled-stack version in the SEO tool stack consolidation breakdown.
Maintenance. Citations decay. Competitors publish. Crawler access regresses silently after a CDN change. A GEO programme that gets funded as a project and not as an operation delivers a good quarter and then quietly reverts, which is a worse outcome than never starting, because it discredits the approach internally.
The one-page format
If you need to walk into a review tomorrow, this is the page:
- The gap, measured. "We are absent from N% of the questions buyers in our category ask an assistant." One number, dated, reproducible.
- The model, with every assumption named and a pessimistic column beside the base case.
- The phases, cheapest first: access audit, retrofit existing pages, entity consistency, then net-new content.
- The three charts you will report monthly, and the date of the first one.
- The kill criterion. "If citation rate is flat at week twelve with access, indexing and rendering verified clean, we stop and reassess." Naming the exit is what makes the entry easy to approve.
That fifth item does more work than the other four combined. Budget owners are not primarily afraid of spending money; they are afraid of open-ended commitments with no defined end.
DigiRank's AI Visibility Tracker produces the measured half of this automatically — citation rate and share of voice per engine, charted over time, alongside your Search Console and GA4 data so the branded-search movement sits on the same timeline as the citations. Tracking starts on the $99/mo Starter plan; the Content ROI engine and visibility timeline that pair citations with published work are included from Agency at $249/mo with a 14-day trial. If you are still deciding whether to run this in-house or hire it out, how to choose a GEO specialist covers the questions worth asking first.
Frequently asked questions
Is generative engine optimization worth the cost compared to traditional SEO? For most businesses it is not an either/or: the work that earns AI citations — clear structure, specific verifiable claims, consistent entity data, third-party corroboration — also strengthens classical search. The genuinely GEO-specific spend is measurement and the access audit, and both are cheap. Treat GEO as a redirection of existing content effort first, and as net-new budget only after you have a measured baseline.
How do I measure GEO ROI when AI answers don't send clicks? Use referral clicks as the floor of the case, not the whole case, and build the model on absence rate, customer value and a deliberately conservative influence rate you label as an assumption. Then watch branded search alongside citation rate — a rising branded-search line during a citation push is the visible fingerprint of the unattributable effect.
What's a realistic influence rate to assume? There is no defensible published figure, so do not quote one as fact. Pick a low number, state it in bold as an assumption, and show the model at that rate and at half of it. If the case only works at the optimistic end, that is a finding, not a presentation problem.
How long before the business case has real data in it? Roughly twelve weeks. Access and indexing verification lands in the first fortnight, first citations on retrofitted pages around week six, and a trend you can defend at about week twelve.
Should labour be in the model? Yes, at a fully-loaded rate. At typical platform prices the hours are the larger line, and a model that omits them loses credibility the moment someone notices — usually mid-meeting.
What should I report to leadership each month? Three charts on one page: citation rate over time per engine, share of voice against named competitors, and measured AI referral plus branded-search movement. Then one paragraph on what changed and what happens next.
How do I compare GEO against spending the same money on ads? On horizon and durability. Ads stop the day you stop paying; a citation earned through depth and corroboration persists across index refreshes and compounds. The fair comparison is cost per influenced customer over a year, not over a month — and it should include the maintenance cost of holding the citations.
What if my competitors are cited and we aren't? That is the strongest possible version of the business case, because the gap is demonstrable rather than theoretical. Show the answer text naming them, then work through why they were selected — see what to do when a competitor is cited in AI answers and you are not.