Does generative engine optimization drive ROI

Generative engine optimization drives ROI when an answer engine cites you on a commercial prompt or stops lying about you on a branded prompt — and you can tie that change to a tagged lead or a dated accuracy incident. It does not drive ROI when a mention dashboard climbs, a vendor GEO score ticks up, or Search Console shows more generative AI impressions on pages nobody buys from. Impressions are exposure. Return is a named opportunity.

This is the spend question under the Answer Engine Optimization playbook. What GEO is lives in GEO explained. How to split a mention from a citation lives in mentions vs citations. This page is the money filter: when the work pays, when it is theater, and what you are allowed to count.

I have been SEO certified since 2021. I will not invent an ROI multiple for this acronym. If a slide needs “2×” or “4.4×” to get approved, the slide is the product.

The short answer

• Pays: cited on “best / vs / hire / price / for [job]” prompts, plus corrections of wrong NAP, offer, or founding facts.
• Does not pay: mention volume on how-tos, glossary pages, and “we got named” screenshots with no URL and no CRM stamp.
• Count: tagged leads (and dated accuracy incidents). Not impressions. Not a GEO score.
• Public numbers you may cite are click, Overview-rate, and answer-share studies. None of them is your return.
• Kill rule: 90 days, frozen commercial panel, eligibility clean, tagged leads still zero → recut or stop. Do not buy a prettier log.

Question	Honest answer
Does GEO drive ROI?	Sometimes. On cited commercial prompts and branded corrections.
Is a mention a return?	No. It is awareness unless a citation or a tagged lead follows.
Is a GSC AI impression a return?	No. It is a supporting-link view in Google’s features.
Is the 2024 GEO paper’s lift ROI?	No. It measured share of generated answer text.
What do I put on the board?	Tagged leads + accuracy incidents, with citation rate as a leading indicator.

If you cannot fill the last row, you are not measuring ROI. You are decorating a retainer.

What counts as ROI for GEO?

A tagged commercial outcome you would have created anyway, plus the cost of the work. Not a visibility percentage. Not “share of AI voice.” Not time-on-site from a Google blog sentence.

Three units survive a finance review. Everything else is a leading indicator or a vanity column.

Unit	What it is	Where it lives	Fake cousin
Tagged lead	Opportunity whose original or assist source is an AI product you can name	CRM, with the prompt or host written down	“They mentioned ChatGPT in the demo” with no field
Accuracy incident closed	Wrong branded fact that stopped appearing on a frozen prompt	Panel log, dated before/after	A press release you published and never re-ran
Protectable conversion	Form, call, or checkout on a URL you can defend as commercial	Analytics → CRM, same opportunity definition	Session conversion on a glossary page

Leading indicators you may keep next to ROI, never as ROI:

Leading indicator	Useful for	Not useful for
Citation rate on commercial prompts	Did the work ship into the answer	Whether anyone bought
Mention rate	Awareness / shortlist presence	Pipeline, unless tagged
GSC generative AI impressions	Google showed a supporting link	Clicks, ChatGPT, Perplexity, closed-won
Brand search	Lagging corroboration	Proof the GEO line paid for itself
Vendor GEO / AI visibility score	Directional log of their corpus	A return you can audit

Formula you are allowed to write on a whiteboard. It is a sketch, not a benchmark:

GEO return ≈ (tagged AI-sourced opportunities × your close rate × your LTV)
             − (hours + tools + invoice)

Accuracy return ≈ (incidents closed × estimated cost of the wrong fact)

If you cannot estimate the first line from CRM, you do not have ROI. You have a measurement quarter. Call it that.

• Opportunity definition matches sales, not marketing’s MQL
• Source field can hold chatgpt, perplexity, overview-assist, or unattributed
• Commercial prompts are tagged separately from vanity brand ego queries
• Cost includes founder hours, not just the agency line
• No multiplier from a blog sits in the cell

Empty boxes are not “directionally positive.” They are empty.

When does GEO spend actually pay?

When the engine names you as a source on a prompt that already sits on the path to a shortlist, or when it stops teaching the wrong story about you. Those are two different jobs. Fund them as two jobs.

Job	Prompt shape	Win	Why money moves
Cited commercial	“best [category] for [job],” “X vs Y,” “hire,” “price,” “agency for”	Named and credited, or at least credited with a URL the buyer can open	Consideration set. The buyer is choosing.
Branded correction	“[Brand] pricing,” “[Brand] location,” “[Brand] vs [competitor]”	Wrong fact gone on the next three panel runs	Trust. A lie in the answer is a lost deal you never see.
Local commercial	“best [service] near me for [job]”	NAP-consistent name + a citeable service page	The click often leaves chat for Maps. Still a shortlist event.

Seer’s April 2026 update — 53 brands, 5.47 million queries, 2.43 billion organic impressions, Jan 2025–Feb 2026 — is the public map of where Google even writes an Overview, not a ROI table. Informational queries showed an AI Overview 36% of the time; commercial 8%; transactional 5%. Inside informational, comparison queries (“X vs Y”) triggered an Overview 95.4% of the time (267 of 280) and question-format 85.9% (1,214 of 1,413) (Seer Interactive).

Read that the operator way: comparison and question pages are where GEO-shaped work can change who gets named. Transactional URLs still live mostly in classic SEO. Do not spend a GEO retainer rewriting checkout copy because an informational how-to got a mention.

Semrush and Kevin Indig’s June 2026 ghost-citation study (3,981 domain appearances, 115 prompts, 14 countries, four engines) split the other axis. Commercial queries (pricing, buying) sat at a 35.6% mention rate and an 84.4% citation rate. Comparative queries produced a 43.3% mention rate — 2.4× the mention rate of informational “what is / explain” prompts, which cited heavily (89.3%) and named brands rarely (18%) (Semrush / Kevin Indig).

So: informational GEO often buys ghosts. Commercial GEO more often buys a source link. Mentions concentrate on comparison. If you only chase mention volume, you will staff the wrong prompts.

1. List the ten prompts sales already hears, or that a lost deal would have asked.
2. Tag each one commercial, branded, or vanity.
3. Run them in ChatGPT with search, Perplexity, and a Google SERP with Overview on/off noted.
4. Fund the commercial misses and the branded lies first.
5. Leave the vanity cluster in a parking lot until the money cluster moves.

A mention on “what is [category]” is a trophy. A citation on “best [category] for [your buyer]” is a bet. Only one of those belongs in an ROI sentence.

Which pages should you fix first if the goal is ROI?

The URLs that can win a commercial or branded prompt, not the URLs that are easiest to mention. Informational volume is a traffic habit. ROI work starts where the engine is asked to choose.

Page job	Prompt it should win	Ship this	Do not
Comparison / vs	“X vs Y for [job]”	Criteria table, named buyer, dated facts	A 2,000-word essay with no grid
Category “best for”	“best [category] for [job]”	Extractable proof: who, where, what you actually ship	A listicle of synonyms
Pricing / packages	“[Brand] pricing,” “what does [category] cost”	Visible numbers or a honest range + what changes the number	“Contact us” as the only sentence
Services / offer	“hire [category] for [job]”	One offer sentence that matches schema and the footer	Three competing taglines
About / entity	Branded who/where/when	Legal name, NAP, founding, sameAs	A manifesto with no facts
How-to / glossary	Informational	Optional. Last.	Your entire GEO quarter

Seer’s comparison-query Overview rate is why the vs page is a commercial object even when Google classifies the query as informational. The buyer is still choosing. Treat that URL as money. Treat the glossary as a parking lot.

Priority procedure:

1. Pull the five URLs that already rank or convert. Those are the candidate set. GEO does not invent a candidate pool.
2. For each, write the one prompt it must win. If you cannot write it, the page has no job.
3. Check extractability: first screen has a sentence and a table a model can lift without the rest of the layout.
4. Check agreement: About, footer, Organization JSON-LD, and Google Business Profile (if you have one) say the same NAP and offer.
5. Only then write new HTML. Publishing a sixth blog without a job is how vanity mentions get born.

Google’s floor for a supporting link in AI Overviews and AI Mode is still ordinary Search eligibility: indexed, snippet-allowed, people-first (Google Search Central: AI features). A beautiful comparison table on a nosnippet URL is unpaid labor.

If you have one week of engineering, spend it on eligibility of the money URLs. If you have one week of writing, spend it on the comparison and offer pages. Blog volume is the leftover, not the program.

How do you run branded correction as an ROI job?

Treat a wrong answer about you as an incident, not a content idea. Correction pays when the lie is something a buyer would use to disqualify you: price, city, service area, “they shut down,” a competitor comparison that cites a dead product.

Wrong thing in the answer	Cost if it stands	Ticket
Wrong city / NAP	Local buyer never calls	Align site, GBP, directories; re-run branded prompts
Wrong price or “starting at”	Qualified lead arrives angry or never arrives	Visible range on the pricing URL; kill stale press quotes
Dead product / old offer	You look like you cannot keep a sentence straight	Update the offer page and the comparison table on the same day
“Closed” / wrong years	Trust tax	About + corroborating third parties that still repeat the error
Competitor owns the vs prompt with your old positioning	You lose the shortlist while ranking	Rewrite the vs page to current facts; ask the citing roundup to update

Procedure:

1. Freeze five branded prompts: who you are, where, price/range, vs the competitor sales actually hears, and “is [Brand] legit / still operating.”
2. Capture the wrong sentence verbatim. Save the URL the engine cited. That URL is the ticket, not your homepage.
3. Fix on-domain facts so the extractable sentence is unambiguous.
4. Fix or request the off-domain source if it still ranks into the answer. One guest post from 2023 will beat your new About page until it does not.
5. Re-run the five prompts on three products. Mark the incident closed only after three consecutive clean runs. One clean screenshot is a fluke.

• The wrong sentence is quoted in the ticket, not paraphrased
• On-domain and the cited third party are both in the ticket
• Close means three clean runs, not a publish date
• You do not call a remaining ghost citation on an unrelated how-to a “brand win”

A closed accuracy incident is ROI-adjacent even with n tagged leads at zero. You stopped paying a tax. Log it as risk avoided, with a note that the dollar value is an estimate from sales, not a dashboard export. If sales cannot estimate it, log the incident anyway. Do not invent a “brand-risk ROI” multiple to make the estimate look scientific.

When is a GEO mention vanity?

When it cannot change a shortlist, a fact, or a tagged lead. Fame without a landing path is a screenshot. Screenshots do not close.

Pattern	What you saw	What it is	What to do
How-to ghost	Your URL in the footnote, brand never said	Citation without memory	Optional. Do not call it pipeline.
Glossary name-drop	Brand in a list of “examples,” no URL	Mention without a path	Log awareness. Do not staff a sprint.
Vendor score up, panel flat	Semrush / Surfer / other GEO grade moved	Their corpus, not your prompts	Ignore until the frozen panel agrees
Competitor roundup	You named, they cited	You lost the click path	Fix extractable proof on the money URL
Ego brand query	“Who is [founder]” with a compliment	Vanity	One accuracy pass. Then stop.
Unattributed “ChatGPT said”	Rep remembered a chat, CRM blank	Anecdote	Add the field. Do not count the story.

The same Semrush study found 61.7% of appearances were ghost citations (source link, no brand name), 25.1% mention without citation, and only 13.2% both cited and mentioned. The full-win cell is the one revenue teams think they bought. If your weekly report celebrates the other 86.8% as “visibility,” you built a vanity machine.

• Every “win” row has mentioned and cited_url as separate columns
• Vanity prompts are labeled vanity before the screenshot
• A mention-only commercial prompt is a ticket (extractable page), not a win
• A ghost on a how-to is not presented as ROI
• Nobody averages informational ghosts with commercial misses into one “AI visibility %”

Awareness is real. It is not a return. Put it in the awareness column and leave it there.

Why don’t impressions count as GEO ROI?

Because an impression is a view, and Google’s dedicated AI report is still a view of a supporting link. Pew already showed how rarely that view becomes a click. Your finance team does not pay rent in impressions.

Pew Research Center (July 22, 2025; March 2025 browsing; 900 U.S. adults; 68,879 Google searches) found traditional-result clicks on 8% of visits when an AI summary appeared versus 15% when it did not. Clicks on links inside the summary happened on 1% of those visits. About 18% of searches in the study produced a summary.

That is a click-rate study on a U.S. panel. It is not your category’s CTR, and it is not a GEO ROI. It is why “we got 40,000 AI impressions” is not a sentence you take to a board without a second sentence about leads.

Google Search Console’s generative AI performance report (announced June 3, 2026) shows impressions: how many times links to your site were shown in AI Overviews and AI Mode. The June launch listed pages, countries, devices, and dates. It was still rolling out to a subset of properties. It does not give you ChatGPT. It does not give you Perplexity. It does not close an opportunity.

Google’s own AI features documentation also claims clicks from pages with AI Overviews tend to be “higher quality,” meaning users are more likely to spend more time on the site. Believe Google for their quality definition. Do not translate time-on-site into revenue. Pew and Google are measuring different objects. Do not pick one and ignore the other.

Metric	What it counts	Allowed in an ROI sentence?
GSC generative AI impressions	Supporting-link views in Overviews / AI Mode	No. Exposure.
GSC Web clicks (mixed)	Blue links + Overview clicks, blended	Not as “GEO ROI.” Too mixed.
Pew 8% / 15% / 1%	Panel click rates, March 2025 U.S.	Cite as pressure. Not your return.
Seer cited vs not-cited CTR	Cohort organic CTR on AIO SERPs	Direction: citation helps vs uncited peers. Still not pipeline.
Tagged CRM lead	An opportunity you defined	Yes.

Seer also reported that being cited in an Overview delivered about +120% more organic clicks per impression versus not being cited on the same AIO-present SERP, and that cited pages still lagged no-AIO pages by about 38% in informational queries. A citation is an advantage on a damaged SERP. It is not a restored 2022 CTR, and it is not a closed-won multiple. Cohort, method, and intent mix are theirs. Copy the shape (split cited / not cited / no Overview). Do not copy the percentage onto your forecast.

Impressions can rise while clicks stay flat. Seer called that out when cited-page impressions doubled and clicks did not. If your GEO report only shows the impression line, you will celebrate coverage you cannot collect.

What did the GEO paper measure, if not ROI?

Share of the generated answer, in a 2023–2024 bench. Aggarwal et al., “GEO: Generative Engine Optimization” (arXiv:2311.09735; KDD 2024) measured Position-Adjusted Word Count (PAWC) — how much of the written answer a page occupied — plus a subjective LLM-judge score. Cite-sources, quotations, and statistics lifted relative PAWC on the order of ~30–40% versus their unoptimized baseline in that setup. Keyword stuffing did not.

That is the origin of half the “GEO ROI” slides on LinkedIn. It is not traffic. It is not leads. It is not 2026 ChatGPT Search. The GEO definition spoke unpacks the bench, the three meanings of the word, and what Google says you can ignore. This page only needs one sentence: do not put PAWC in a return formula.

Object	Paper	Your P&L
Win metric	PAWC / impression-of-answer	Tagged lead or closed incident
Engine	Lab retrieve-then-write + a then-current Perplexity slice	Live products your buyers use
Edit that helped	Evidence, quotes, stats on the page	Same direction, still not a multiple
Edit that failed	Keyword stuffing	Still fail. Still not ROI.

Direction you may keep: evidence-bearing pages occupy more of a synthesized answer than padded ones. Number you may not keep: “+40% GEO ROI.” That sentence is a category error.

If an agency quotes the paper as a pipeline forecast, ask them which column in GEO-bench maps to your CRM. There is not one.

Why you should refuse a GEO ROI multiple

Because the published conversion premiums do not agree, and none of them was measured on your funnel. The range in circulation in 2025–2026 runs from a modest gap to carnival numbers. Treating any of them as a planning constant is how you invent a return.

Claim you will hear	What to do	Why
“AI leads convert 4.4×”	Reject unless you can read sample, definition of conversion, and whether modeling was used	Circulates as Semrush-attributed marketing; not your close rate
“ChatGPT converts 7% vs Google 5%”	Treat as a gap in someone else’s transactional cut, not a law	Similarweb-style referral studies are not a GEO multiplier
“AI traffic converts ~1.3× organic”	Still not yours	Different sites, different conversion events (signup vs purchase vs form)
“GEO ROI = (AI pipeline − spend) / spend”	Fine as algebra. Useless without tagged pipeline	The numerator is the whole argument
“Our GEO score went from 42 to 67”	Not ROI	Proprietary corpus

I will not pick a winner among those multiples. I will not average them. I will not “be conservative” by using 1.5× so the model feels responsible. If you need a number to approve the work, use your close rate on your tagged sample, published as a fraction (3/18 vs 40/200), and refuse a headline percentage until one extra win cannot flip the story.

Google saying Overview clicks are higher quality (more time on site) is a quality claim. Time is not revenue. A longer session on a blog post you cannot convert is not ROI. A shorter session on a pricing page that becomes an opportunity is.

• The model’s conversion rate is yours, from CRM
• The model’s volume is tagged, not “AI-influenced” by vibe
• Cost includes internal hours
• The paper’s PAWC lift is not in the spreadsheet
• The slide title does not contain “×”

If those boxes fail, you are not forecasting. You are fundraising.

How do you measure GEO with tagged leads?

Stamp the opportunity, then compare fractions. Referrer hostnames, UTMs on links you control, and a self-report field. Overview clicks still look like google.com. Do not “fix” that with a regex.

This is the instrumentation, not a close-rate bake-off against organic. The bake-off is a different question. Here you only need enough tagging to answer: did the GEO line produce commercial opportunities we can name?

Source you can see	Put in CRM as	Do not
chatgpt.com / chat.openai.com	ai:chatgpt	Fold into organic
perplexity.ai	ai:perplexity	Call it “referral” and forget
Claude / Gemini app hosts you actually observe	ai:[host]	Invent a host you have not seen
UTM you set on a citeable URL (utm_medium=ai)	Same as the host you intended	Tag every page “just in case” and pollute the set
google.com (including Overview)	Organic. Optional note: overview-possible	Split with a guess
Direct / copy-paste / in-app, no Referer	Unattributed, unless self-report is specific	Dump into AI to enlarge the sample

Procedure:

1. Freeze 25–40 prompts. At least half commercial or branded-correction. Write the IDs down. Do not rotate them weekly to chase a better screenshot.
2. Run the panel weekly in the products your buyers use. Log mentioned / cited URL / absent / wrong.
3. In analytics, allow-list the AI hosts you have actually seen. Create a GA4 channel or a CRM pick-list that matches.
4. On every new opportunity, require original source. If sales skips it, the week does not count as a measurement week.
5. Monthly, publish: citation rate on the commercial slice, tagged AI opportunities (n), unattributed (n), cost (hours + invoice).
6. Do not compute ROI until n on the tagged side is large enough that one extra win is not the whole story. Until then, report the fraction and the panel.

Self-report (“ChatGPT recommended you”) is an assist, not a last click. Keep it. Do not let it overwrite a clean google.com last touch. Three disagreeing fields mean you do not have one GEO close rate. You have a triangulation problem. That is still better than a invented multiple.

• Prompt IDs frozen for 90 days
• Commercial slice reported separately
• Host allow-list reviewed against real logs, not a blog
• Overview stays inside organic
• Monthly readout is a table, not a sparkline of impressions

A tagged lead with a cited commercial prompt in the same week is the strongest story you can tell without lying. It is still not a causal proof. It is a defensible correlation. Take it. Do not dress it up as a randomized trial.

What does a finance-readable GEO log look like?

One row per prompt-run, plus a monthly rollup that finance can read without a GEO glossary. If a CFO needs a decoder ring, you built a marketing log.

Prompt-run columns (copy these; do not add a “visibility %”):

Column	Allowed values	Ban
prompt_id	Frozen ID	A newly worded prompt “because it worked”
intent	commercial / branded / vanity	Blank
product	chatgpt-search / perplexity / overview / other named	“AI”
mentioned	Y/N	“kind of”
cited_url	URL or —	“they used us”
accuracy	ok / wrong / n/a	“mostly fine”
wrong_span	Quoted sentence or —	A vibe
date	ISO date	“this week”

Monthly rollup:

Line	How to fill it
Commercial citation rate	Cited commercial runs ÷ commercial runs. Show a/b, then a percentage if you must.
Branded accuracy	Clean branded runs ÷ branded runs. Same fraction rule.
Vanity mentions	Count, in a footnote. Not in the ROI sentence.
Tagged AI opportunities	CRM count for the month. Fraction of all new opportunities if useful.
Unattributed	Direct + self-report-only. Keep visible so you cannot hide.
GSC gen-AI impressions	Optional appendix. Not a rollup line that leads the email.
Cost	Hours × loaded rate + tools + invoice.

Sample month that is allowed to go to finance (illustrative structure, not a client result):

Line	Example shape
Commercial citation	4/20 on the frozen slice
Branded accuracy	8/10 (two NAP leftovers; tickets open)
Tagged AI opps	2 (chatgpt.com ×1, perplexity.ai ×1)
Unattributed	11
Cost	Hours + invoice, one number
Claim in the email	“Two tagged opps, citation 4/20, two branded tickets still open.”
Claim you do not send	“GEO ROI is 3.2× this month.”

If tagged opps are 0/0 because the CRM field did not exist, you failed measurement, not GEO. Fix the field. Do not infer ROI from the citation rate alone.

• Rollup fits on one screen
• Fractions appear before percentages
• Vanity is footnoted
• Cost is a number, not “covered by SEO”
• No multiplier cell

A log like that can survive a skeptical operator. A GEO score cannot.

What usually fails first?

The scoreboard collapses into mentions, then the work chases the scoreboard. Eligibility and entity facts are the boring failures. Vanity measurement is the expensive one.

Failure	What it costs	What you do instead
Mention vanity report	You staff how-tos that get named and never cited on money prompts	Split mention vs citation. Tag prompts commercial / vanity.
Impression theater	GSC AI impressions up, CRM silent, retainer renewed	Stop putting the impression line in the ROI slide.
Paper-as-forecast	“+40% visibility” sold as pipeline	Cite PAWC as historical answer-share. Cut it from the model.
Multiplier shopping	Budget approved on 4.4×, missed on actual n	Fractions only. Kill rule on tagged leads.
Eligibility skip	Unindexed or nosnippet money URLs; Google cannot cite what it cannot snippet	AI features: indexed + snippet-eligible. No extra Overview markup.
Wrong-fact leftover	Site fixed, Perplexity still cites an old roundup	Corroboration ticket. Re-run branded prompts. Log the incident closed only after three clean runs.
No owner	Panel dies in week three; screenshots from sales calls become the dataset	Name one person. If nobody has two hours a week, do not buy a program.

Concrete pattern from audits: a studio ranks for its category, ChatGPT recommends three competitors with clearer About pages and comparison tables, and the monthly GEO deck shows a rising mention series on “what is [category]” posts. The commercial prompts never moved. The invoice did.

Wrong-fact leftover is the other expensive miss. A founder corrects a price or a city on the site. The branded prompt still recites the old number because a directory and a 2023 guest post agree with each other. You did not fail to “do GEO.” You failed to treat correction as a corroboration job. Silence on a branded lie is not neutrality. It is a sales tax.

If the panel is not frozen, every week is a new anecdote. Anecdotes renew retainers. They do not measure return.

How long until GEO shows results you can defend?

A baseline exists in week one. Citation movement, if it comes, is a 30–90 day object. Tagged-lead ROI is later than that, and it may stay at zero. Anyone selling Tuesday is selling a screenshot.

Horizon	What you can honestly have	What you cannot
Week 1	Frozen panel, eligibility pass, commercial vs vanity tags, CRM field	A return
Days 14–30	First re-runs; maybe a branded fact that was already almost right	A stable citation rate
Days 30–90	Citation / accuracy deltas on the frozen set; first tagged hosts if volume exists	A multiple
After 90	Keep, recut, or kill using tagged n + commercial citation rate	“The GEO score proves ROI”

Google still recrawls on its schedule. ChatGPT Search and Perplexity are not obligated to notice your publish. Off-site corroboration lags on-site HTML. A 90-day loop is the honest unit because non-determinism eats single screenshots. The playbook’s measurement layer is built for that. This page only adds the kill criterion: commercial citation rate and tagged leads, not impressions.

If week 12 is flat on commercial prompts and eligibility is clean, you do not have a “needs more content” story by default. You have a page-job or corroboration miss, or you are on prompts engines do not answer with vendors. Recut the pages. Do not recut the acronym.

Local and high-LTV services can see a tagged call sooner because the path is short. Low-volume B2B with a six-month cycle will not. Do not compare those two clocks. Compare each program to its own n.

What should you skip if you only have a week?

Skip the content calendar, the tool bake-off, and the ROI model that needs a borrowed multiple. A week can produce a money-prompt list and a tagging rule. It cannot produce return.

Do this week	Skip this week
Write 15 commercial + 5 branded prompts from sales language	40 informational posts “for GEO”
Run them once. Log mention / citation / wrong / absent	A vendor GEO trial as the scoreboard
Inspect money URLs: indexed, snippet-eligible, extractable table	Sitewide schema theater
Add CRM source values for AI hosts you have seen	A last-click GEO ROAS cell
Circle two URLs that should win a “best / vs” prompt	Rewriting the homepage into a glossary
Write the kill rule: 90 days, tagged n, commercial citation rate	A 12-month retainer “to not fall behind”

One-week procedure:

1. Export the last 20 closed-won and 20 lost notes. Highlight any AI product named. If none, you still might need a panel — buyers do not always confess — but you do not have qualitative proof yet.
2. Turn those jobs into prompts. Freeze the list.
3. Screenshot three products. Do not optimize the prompt mid-flight to get a prettier cell.
4. Tag two URLs as the commercial bets. Everything else waits.
5. Create the CRM field even if n will be zero this month. Zero is a number. Missing is a dodge.

A week of instrumentation beats a quarter of mention screenshots. Bravery is not a measurement strategy.

When is GEO not worth doing yet?

When you cannot staff a panel, when eligibility is on fire, or when nobody in the category asks an engine to shortlist a vendor. Education can still be cheap. A retainer cannot.

Condition	Verdict	First move
Crawl, noindex, or nosnippet on money URLs	Wait	Technical SEO floor. Google’s AI features still require indexed, snippet-eligible pages.
NAP / offer / legal name disagree across About, footer, schema	Wait on expansion	Truth layer. Branded prompts will keep lying.
No owner for two hours a week	Do not buy a program	DIY ten prompts a month, or nothing.
Category is relationship-only; buyers never research	Light or skip	Re-check quarterly. Do not invent demand.
LTV cannot absorb an audit plus a month of fixes	Cap hard	Entity packet + three answer-first service pages. No retainer.
You want GEO to reverse an informational traffic cliff	Wrong tool	Zero-click compression is a product change. Citation work does not restore 2022 CTR.
The only KPI on offer is a GEO score or GSC AI impressions	Do not buy	Change the scoreboard or walk.

Waiting is not “ignoring AI.” It is refusing to fund theater on a broken site. The worth-it gate for AEO as a budget object is a cousin of this page; the difference here is the unit: tagged leads and branded corrections, not a generic “is AEO worth it.”

If sales has never heard ChatGPT or Perplexity and your commercial prompts come back empty for every vendor, you may be on a surface engines do not use for shortlisting. Log that. It is a finding. It is not a reason to buy a larger dashboard.

Fund GEO when the commercial panel already names competitors and not you, or when branded answers are wrong, and someone will log it. Otherwise you are buying a noun.

How do you cost a GEO quarter without fake ROAS?

Add hours, tools, and invoice. Compare that sum to one incremental closed job at your real close rate. If you cannot name the job, you are buying a measurement quarter. Cap it like education.

Cost line	Include	Easy dodge
Internal hours	Panel owner + writer + whoever ships HTML	“We were going to blog anyway”
Tools	Prompt-log seat, Search Console (free), CRM field (free)	Double-counting Semrush you already pay for SEO
Invoice	Audit or execution, whatever you actually paid	Bundling it into “retainer” so it disappears
Opportunity cost	Money URLs you did not improve	Pretending blog volume is free

Decision list, not a calculator you can screenshot into a pitch deck:

1. Write the all-in cost for 90 days.
2. Write your close rate and average closed value from finance, not from a GEO blog.
3. Write how many tagged opportunities you would need at that close rate to cover the cost. That is a coverage count, not a forecast.
4. If the coverage count is “one extra job” and your LTV supports it, an audit-plus-90 is rational even with n currently at zero — you are buying a test with a cap.
5. If the coverage count is “a dozen extra jobs you have never seen,” you are not buying GEO. You are buying a hope. Shrink the scope.

I will not fill those cells with Spurlock Studios fees or a made-up win rate. The object is the arithmetic. The numbers are yours.

Test result at day 90	Move
Commercial citation up, tagged n still 0, field was live	Keep the panel; recut pages; do not scale spend
Tagged n covers the coverage count	Keep, still without a multiple
Panel flat, eligibility dirty	You never ran GEO. Run eligibility.
Panel flat, eligibility clean, competitors also absent	Category may not shortlist in these products. Cap.
Mentions up, commercial citations flat	Vanity. Stop staffing the glossary.

A coverage count is honest because it does not pretend the numerator arrived. A ROAS cell on citations is dishonest because Pew’s 1% in-summary click rate should humble anyone who last-clicks an Overview into a return.

FAQ

Does generative engine optimization drive ROI?

Sometimes. It drives return when you get cited on commercial prompts or you close a branded accuracy incident, and you can stamp that work onto a tagged lead or a dated log. Mention volume, vendor GEO scores, and Search Console generative AI impressions are exposure. They are not ROI. I will not invent a multiple that makes the “yes” easier.

How do I measure whether does generative engine optimization drive ROI is working?

Tag CRM opportunities with AI hosts you actually see, keep Overview clicks inside organic, and report commercial citation rate next to tagged n — as a fraction, not a headline percentage. Re-run a frozen prompt panel weekly so inclusion is a time series. If you cannot name the source field and the commercial slice, you are not measuring whether it is working. You are watching impressions.

What usually fails first when teams try this?

The scoreboard. Teams merge mentions and citations, then staff the prompts that photograph well. Eligibility misses (nosnippet, unindexed money URLs) and leftover wrong facts off-site are the next failures. A rising GEO grade with a flat commercial panel is the tell. Fix the log before you buy more posts.

How long does this take to show results?

A baseline is a week. Citation and accuracy movement, if any, is a 30–90 day object on a frozen panel. Tagged-lead ROI is slower and can stay at zero, especially on long B2B cycles. Anyone promising a Tuesday screenshot as return is selling the screenshot. Set a 90-day kill-or-recut date before you start.

What should I skip if I only have a week?

Skip the editorial calendar, the tool bake-off, and any model that needs a borrowed conversion multiple. Write commercial and branded prompts from sales language, run them once, inspect two money URLs for eligibility, and add the CRM source field. A week of tagging beats a quarter of vanity mentions.

When is this not worth doing yet?

When money URLs are not eligible to be cited, entity facts still conflict, or nobody will own a two-hour weekly panel. Also skip a retainer if the proposed KPI is impressions or a vendor score, or if you are trying to restore 2022 informational CTR. Cap to a DIY panel until those gates are honest.

CTA

Stop treating mention dashboards as return. Freeze the commercial prompts, tag the leads, and kill the work if n stays at zero.

Lane: /visibility · Book a visibility audit.