How to prove GEO ROI and AI-SEO ROI to clients when attribution is hard: the first-party metrics that hold up, why share of voice is weak, and a reusable before/after framework.
The hardest part of selling GEO is not doing the work. It is the review meeting three months later when a client asks what the AI-SEO retainer actually bought them. You know the work mattered. You added llms.txt, cleaned up the content assistants read, fixed the technical gaps that kept AI bots out. But when you open analytics, the win that should be obvious is hiding inside a "(direct)" bucket with no referrer attached, and the number your monitoring tool puts front and center is a share-of-voice percentage you cannot fully explain. This post is the framework I reach for to turn that into a client-ready story that holds up to scrutiny.
Quick orientation before the detail. Crawlytics is an AI-agent-readiness platform, and it does not do prompt-sampled share-of-voice monitoring, so I am not going to tell you that metric is the answer. The case here is the opposite: lead with first-party signals you can audit, and treat the vendor-sampled numbers as context. I will be honest about where attribution genuinely breaks, because pretending it does not is how agencies lose accounts when a client checks the math.
GEO ROI is hard to prove because the entire measurement chain that traditional SEO relies on is missing a link. In classic search, someone queries Google, clicks a blue link, and arrives with a referrer that tells you exactly where they came from. You count the session, follow it to a conversion, and the ROI story writes itself. AI search broke that chain in three specific places.
Assistants strip the referrer. When ChatGPT or a similar assistant recommends a client and the user clicks through or types the brand into a browser, the session frequently arrives with no referrer at all. Your analytics files it under "(direct)" or "(none)," the same bucket as bookmarks and typed URLs. The most valuable GEO outcome, an AI recommending you, is the one your tools are worst at labeling. We wrote a full walkthrough of the ChatGPT "(direct)" traffic problem and how to recover it, and it is the single biggest reason GEO looks like it is doing nothing when it is working.
The funnel is shorter and fuzzier. AI answers compress the research-to-decision path. A user can read a synthesized answer that cites your client, form an opinion, and never visit the site until they are ready to buy, if they visit at all. The middle of the funnel that you used to measure as on-site sessions happens inside someone else's chat window. We unpacked this shift in how AI search changes the SEO funnel. Fewer measured touches does not mean less influence. It means your influence is harder to instrument.
Sampling is not the same as measurement. The tools that promise to fix this mostly do it by sampling: they ask a fixed list of prompts across a few LLMs on a schedule and record whether the brand appeared. That is a reasonable signal, but it is an estimate of perceived presence, not a count of real demand. LLM outputs vary between sessions, so the same prompt can return a different answer tomorrow with nothing changed on the client's site.
Put those three together and you get the core honesty problem. The metric clients most want, a clean "AI sent us X visitors who converted Y times," is the one nobody can produce exactly today. Any agency that hands a client a precise AI-attribution figure is either guessing or hiding the asterisks. The way to keep an account is to be the one who says so, then shows the numbers that do hold up.
Five metrics survive a skeptical client review. Ranked roughly by how defensible they are, from most to least.
1. Server-log AI-bot traffic growth. This is the most defensible GEO metric you have, because it comes from your own logs and cannot be sampled, guessed, or inflated. When GPTBot, ChatGPT-User, ClaudeBot, PerplexityBot, and the rest crawl a client's pages, those hits are real first-party events. A rising trend after you ship llms.txt and clean content shows AI systems are doing more work to ingest the site, which is the precondition for everything downstream. Crawlytics reads this directly from server-side logs across 25+ bot signatures and 19 providers, so you can show a client a real before-and-after crawl curve rather than a vendor's estimate. If you are setting this up, our guide on tracking AI activity covers the log side.
2. AI referral traffic recovered from "(direct)". Some assistants and AI surfaces do pass a referrer, and those sessions can be isolated by referrer host. The bigger move is recovering the AI-influenced slice hiding in "(direct)": tag every link you control with UTMs, watch for AI referrer patterns, and track whether the no-referrer segment grows in step with your GEO work. This is best-effort, not exact. Assistants strip parameters, and you will never recover all of it. But a "(direct)" line that climbs alongside rising bot traffic is a corroborating signal a client can follow.
3. Citation and mention trend. If the client engagement includes a prompt-sampling monitor, the trend in how often the brand is cited or mentioned in AI answers is worth reporting, as a trend. Whether the brand shows up for the queries that matter, and whether that share is moving the right direction over time, is a legitimate input. Just hold it at the right altitude: it is sampled, it is directional, and it belongs in the body of the report, not the headline. Crawlytics does not generate this number, so if a client needs deep multi-prompt mention tracking, pair the readiness work with a dedicated monitor and present both honestly.
4. Agent-readiness score, before and after. A readiness score is the cleanest before/after artifact in the whole framework, because it is deterministic and you control the timing. Run a baseline scan at kickoff, ship your work, run it again. The free Crawlytics agent-readiness grader scores five categories on a live scan: Discoverability, Content Accessibility, Bot Access Control, Protocol Discovery, and Attribution. A jump from, say, a D to an A on a dated, itemized scan is the kind of concrete artifact a client can hold, and unlike a traffic number it is not at the mercy of seasonality or someone else's algorithm.
5. AI-driven conversions. Where a client has a lead form or a checkout, you can attribute conversions on the same best-effort basis as referrals: UTM-tagged links, AI referrer hosts, and the "(direct)" segment reviewed for correlated spikes. For ecommerce specifically, an agent-commerce path can observe conversions from agent-driven sessions client-side. Crawlytics's WebMCP commerce layer records conversion attribution across five payment providers for sites that run it, and you can see the live version on staystrat.com. Frame this as experimental and directional. The honest label is "AI-influenced," not "AI-caused." More on the mechanics in our AI attribution feature page.
Do not make prompt-sampled share of voice your headline ROI number. It is the metric most monitoring vendors push hardest, and it is the one most likely to embarrass you in a client meeting, for one structural reason: the denominator is unauditable.
Share of voice is "your brand's appearances divided by total appearances across a set of prompts." The vendor chooses that prompt set, and you usually cannot inspect it, change it, or verify it represents real demand. So when the number moves, you cannot tell a client why. Did the brand actually gain ground, or did the vendor add prompts, or did the underlying models update? You are reporting a percentage with a moving, hidden bottom half. We laid out the full case in why AI share of voice is a made-up number, and the short version is that it is fine as a directional read and dangerous as proof.
There is a second trap. Share of voice is a relative metric, so a client can "win" it while their actual AI traffic falls, simply because competitors fell faster. A number that can improve while business outcomes decline is not an ROI metric. It is a vanity metric wearing an ROI costume. Keep it in the report as context, label it as sampled and directional, and never let it carry the weight of justifying the spend on its own.
A before/after story works when you captured the "before." The most common reason agencies cannot prove GEO ROI is that nobody ran a baseline, so there is nothing to compare the "after" against. Fix that at kickoff, every time.
At engagement start, capture four things and date them. Run the agent-readiness scan and save the score plus the itemized findings. Pull a clean read of current AI-bot crawl volume from the logs. Snapshot AI referral and "(direct)" traffic for the prior period. If a monitor is in play, record the starting citation trend. That dated bundle is your baseline. Without it, every later number is a figure floating in space.
Set the time expectation in the same conversation. Readiness score and bot-crawl activity can move within weeks of shipping llms.txt and clean content. Referral and citation changes lag, because models re-crawl and re-index on their own clock. Telling a client at kickoff that the first review is a baseline read, not a verdict, prevents the "why is nothing happening" panic at week six.
Then report the deltas, paired. One metric alone is easy to dismiss. Bot traffic up 60%, readiness score D to A, "(direct)" sessions up in the same window, and three new citations is a pattern, and patterns persuade. The art is showing several independent first-party signals all pointing the same way, then naming the specific work that produced them: the llms.txt you generated, the content you made readable to bots, the access blocks you removed.
Here is the framework I reuse across accounts. Five steps, designed to be defensible.
One disclosure to put in writing inside the report itself: AI attribution is best-effort. Assistants strip referrers and UTMs, so referral and conversion figures are directional, not exact last-click counts. The first-party metrics, bot traffic and readiness score, are the audited backbone. Saying this plainly does not weaken the report. It is what makes a sophisticated client trust the parts that are solid. If you run AI-SEO for multiple brands, the same framework scales, and we cover the operational side in AI visibility tools for agencies and how to package it in white-label AI search reports.
Written by Crawlytics Team. Crawlytics tracks AI bots, generates llms.txt, and powers WebMCP commerce, all from one snippet on any stack. See how it works →
Measure GEO ROI with first-party signals you can audit: server-log AI-bot traffic growth, AI referral sessions recovered from "(direct)" via UTMs and referrer patterns, agent-readiness score before and after your work, and AI-attributed conversions where a lead or checkout path exists. Set a baseline window before the work starts, then report the deltas. Treat prompt-sampled share of voice as supporting context, not the headline, because its denominator is not auditable.
The metrics that prove AI-SEO is working are rising AI-bot crawl volume in your server logs, growth in AI referral traffic from ChatGPT, Perplexity, and similar sources, an improving agent-readiness score, more citations or brand mentions in AI answers if you run a monitor, and AI-influenced conversions. Bot traffic and readiness score are the most defensible because they come from your own logs and your own pages, not a third-party sample.
Share of voice is useful context but a weak headline ROI metric. It is built on a prompt list the vendor chooses and you cannot inspect, the denominator is unstable, and LLM answers vary between sessions, so the percentage can swing without anything changing on your site. Report it as a directional trend alongside first-party metrics, and never present it as audited proof of business impact.
Track AI-driven conversions by tagging the links you control with UTMs, watching for AI referrer hosts in analytics, and reviewing the "(direct)" / no-referrer segment for spikes that correlate with your AI-readiness work, since assistants often strip referrers. Where a lead form or checkout exists, attribute conversions on the same best-effort basis. Label all of it experimental and best-effort. It is a strong directional signal, not exact last-click attribution.
Expect early signals in 30 to 90 days and a clearer trend by one to two quarters. AI-bot crawl activity and agent-readiness score can move within weeks of shipping llms.txt and clean content. Citation and referral changes lag, because models re-crawl and re-index on their own schedule. Set this expectation with the client at kickoff so the first review is framed as a baseline read, not a verdict.
This page is part of Crawlytics.app. View all pages: llms.txt · llms-full.txt