How to Measure AEO Agency ROI: Metrics, Attribution, and a 12-Month Timeline (2026)

How to Measure AEO Agency ROI

Most AEO reporting is theater. An agency shows you a rising "AI visibility" line, you nod, and nobody in the room can say whether that line put a single dollar into pipeline. If you are paying for answer engine optimization, you need a way to tell real return from a pretty chart.

TL;DR. Measure AEO ROI across four tiers instead of one: crawler activity, share of answer, AI-sourced traffic, and pipeline revenue. Visibility without pipeline is a vanity number. Expect first citations in weeks and revenue attribution in 8 to 12 months. No public B2B SaaS conversion benchmark exists yet, so you measure your own.

The four-tier AEO measurement ladder

Every credible page on this topic, and every AI engine we asked to answer it, converges on the same shape: a ladder from leading indicators to business outcomes. Copy this table into your next agency reporting review and make them fill every row.

The tiers are not interchangeable. Tier one tells you the engines noticed your content. Tier four tells you the work paid for itself. An agency that reports only tier two is showing you the middle of the ladder and hoping you never ask about the top or the bottom. That is the single most common way AEO reporting flatters itself.

A real AEO agency reports the whole ladder, and visibility is only the middle rung

Ask three answer engines how to judge an AEO agency and they all say the same thing without being prompted: report visibility, then traffic, then engagement, then revenue, and connect them. An agency that hands you an "AI visibility" percentage with nothing tied to leads or pipeline has given you the least useful number it owns.

Start with the definitions, because most disputes about ROI are really disputes about what a word means.

Tier one, the leading indicator. Before an engine cites you, its crawler fetches you. That gap between fetching and citing is itself a signal. Cloudflare defines the crawl-to-refer ratio as the AI-bot HTML fetches divided by the human referrals that platform sends back. In the week of June 19 to 26, 2025, Cloudflare measured Anthropic's crawler pulling "nearly 71,000 HTML page requests for every HTML page referral." Read that as a warning about the top of the ladder: bots extract far more than they send back, so raw crawl volume is a leading indicator of interest rather than a promise of traffic. Server logs are the highest-fidelity version of this signal because they record what actually hit your origin instead of an outside estimate. If your agency has never asked for log access, ask why.

Tier two, visibility. Peec, the tracker we run internally, defines visibility as the "Percentage of AI responses where your brand appears." Its formula is plain: responses that mention your brand, divided by total responses, times 100. Alongside it sit citation rate (how often your URL is referenced in the answer text) and position (your average rank when you show up, where lower is better). These last two are averages, and citation rate can exceed 1.0, so do not read them as percentages. This is where most reporting lives, and where most of it also stops.

Tier three, AI-sourced traffic. The count of real humans arriving from an AI answer, plus what they do next. This is where measurement gets slippery, and where most agencies quietly go silent. The slipperiness is not their fault. A large share of these visits shed their referrer on the way in and get miscounted, which is the attribution problem covered further down. A capable agency treats that gap as a known tax and corrects for it. A weak one just reports the undercounted number and lets you assume it is the whole story.

Tier four, pipeline and revenue. The only tier that survives a CFO's questions: attributed opportunities, closed revenue, and whether the spend returned more than it cost.

When we ran our own AEO program on ourselves, the four tiers moved in sequence, not together. In one quarter, LoudFace went from appearing in 0.18% of the AI answers in our category to 10.35%. Brand mentions climbed from 8 to 1,184 in a 30-day window, average citation position settled at 2.9, and AI-referred traffic reached 4.75% of pageviews over the last 28 days. Visibility moved first. Traffic followed. That ordering is the whole point of the ladder, and the full breakdown lives in our own AEO case study.

The ROI formula, and the problem with the conversion data

The formula is not complicated:

ROI = (attributed revenue − agency cost) / agency cost × 100

A worked example, with illustrative round numbers that are not a LoudFace rate. Say attribution ties two dollars of closed revenue back to AI-sourced discovery for every one dollar the program costs. ROI = (2 − 1) / 1 × 100 = 100%, double the spend returned. The arithmetic is trivial. The hard part is the word "attributed" sitting in the numerator, which is the attribution problem the next section takes apart. Get that number honest and the ROI takes care of itself. Fudge it and the percentage is fiction.

One caveat before you trust the conversion numbers below. You will see impressive AI conversion stats quoted as if they settle the ROI question. They do not, because none of the large public numbers are about B2B SaaS.

Adobe Analytics, working from more than a trillion visits to US retail sites, found AI traffic "converted 42% better" than other visitors in March 2026. That is a real, large dataset. It is also retail. A year earlier the same metric ran 38% worse, and in July 2024 it was 43% worse, so the reading is a maturation curve (from minus 43% in July 2024 to minus 38% in March 2025 to plus 42% in March 2026) rather than a fixed truth. Similarweb, measuring across the whole web rather than one vertical, reports ChatGPT referral traffic converting at 7.1%, second only to paid search at 7.8%. Different method, different population, and you cannot average the two numbers into one.

There is no published B2B SaaS AI-referral conversion benchmark. The retail and cross-vertical figures tell you the channel is maturing fast. They tell you nothing about your funnel. So the only conversion rate that matters for your ROI is the one your own CRM produces, which is why tier four depends on attribution you build rather than a benchmark you borrow. For how spend maps to these tiers, see our breakdown of what AEO retainers actually buy.

Visibility is not share of voice, and mentions are not citations

Two conflations wreck more AEO reports than anything else, and both are easy to weaponize into a rosier number than reality.

Visibility versus share of voice. These come off the same data and produce different figures. Peec's own worked example: a brand mentioned 4 times across 10 chats reads 40% visibility, but if a competitor was mentioned 12 times, the same brand's share of voice is 25% (4 divided by 16). Visibility is how often you appear. Share of voice is how much of the conversation you own against everyone else. Share of voice is always the lower, harder number when competitors are in the room. An agency reporting the 40% and calling it dominance is handing you the flattering half of one calculation.

This is why we label the headline metric visibility or share of answer, and never share of voice. When Toku "owns 86% visibility on the core crypto-payroll prompt" at position 2.4, that 86% is visibility, meaning it appears in most answers to that question. Its overall share of voice across the twelve tracked brands sits near 8%. Both are true. They measure different things, and a report that swaps one for the other is lying by relabeling. Toku's numbers come from a sample of 75 prompts, 12 brands, and 6,868 AI responses, detailed in the Toku case study.

Mentions versus citations. A mention names your brand in the answer. A citation links your URL as a source. Peec calls these brand visibility and source visibility, and they move independently. You can be cited without being named, and named without being cited. The first is an awareness signal. The second is the referral signal that actually sends traffic. Reporting mention volume as if it were citation performance overstates ROI, because mentions do not produce clicks and citations sometimes do.

One more trap: the blended number. Per-engine standings diverge hard. A single blended visibility figure averages a winning engine and a losing one into a number that hides both. You might be dominating Perplexity and invisible on ChatGPT, and the blended average will read as a comfortable middle that describes neither. The two engines reward different things and need different work. A blended number buries which lever to pull. Ask for the split. Every time, ask for the split.

Attribution: how to actually track AI visits, and what GA4 misses

Here is the uncomfortable floor. A large share of AI-referred visits arrive with no referrer header and land in your "Direct" bucket, so raw channel reports undercount AI by a wide margin. Practitioner estimates put 35% to 70% of AI-referral sessions in Direct. You cannot fix a number you cannot see, so attribution is a stack rather than a single tool.

Build it in four layers, and stand the whole thing up on day one so your baseline is honest from the start.

GA4's native AI Assistants channel. Google now ships an "AI Assistants" default channel, defined as the path by which "users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok." Useful, and free, with two named gaps: it does not list Perplexity or Claude, and it explicitly excludes Google's own AI Overviews and AI Mode, which stay filed under organic search.

A custom channel group. A regex-based channel group catches the referrers Google's native list misses. Perplexity reliably passes a perplexity.ai referrer across desktop and mobile, so it is recoverable this way. ChatGPT only began passing UTM parameters in June 2025 and still drops attribution from its mobile app, so expect leakage there.

Self-reported attribution. A "how did you hear about us" field on your demo or contact form is the practitioner backstop for everything the referrer loses to Direct. It is the one place you capture an AI-discovery touch that no channel report will ever show you. For a B2B SaaS buyer, one honest form answer often beats a week of dashboard archaeology.

Server-log crawler analysis. Your own logs record which AI bots fetched which pages, deterministically, from inside your origin. This is the tier-one leading indicator, and it is higher fidelity than any external probe because it is not an estimate at all.

Every engine has a measurement blind spot

No tool sees all of AI search, and honest reporting names the gaps instead of papering over them.

• Native apps send no referrer. Traffic from the ChatGPT and Claude desktop and mobile apps frequently arrives with no referrer header. This is the structural floor every referrer-based tool hits, GA4 and Cloudflare's ratios included. It is a law of the plumbing. It means every AI traffic count you have is a floor rather than a ceiling.
• Google AI Overviews and AI Mode are not separable in GA4. Clicks from Google's AI surfaces stay inside organic search natively, so you cannot cleanly isolate them without extra tooling.
• Perplexity and Claude are missing from GA4's native list. Perplexity you recover through its referrer. Claude native-app traffic you largely do not.
• Ahrefs Brand Radar tracks Claude with custom prompts only. Claude visibility is not in its default cross-engine pull, so a Claude reading there takes deliberate setup.

The takeaway for a buyer: when an agency reports one blended AI number with no caveats, that is a tell. The people who actually measure this know where their instruments go blind, and they tell you before you have to ask.

The 12-month timeline: when each tier actually moves

Nobody has published an external, peer-reviewed "time to first AI citation" benchmark. Anyone who quotes you a precise industry-standard timeline is inventing it. What follows is our first-party model, earned by running this loop on our own site and our clients', and it maps cleanly onto the four tiers.

• Month 0 to 1. Baseline everything. Stand up the attribution stack, snapshot current visibility per engine, and start reading server logs. The fastest AI surface, Google AI Overviews, sits on the live index and can pick up new content within days, so early crawler activity is what you watch first. More on why the surfaces move at different speeds in how long AI citations take.
• Month 2 to 4. Visibility starts moving. Citations consolidate into a repeatable slot, share of answer ticks up on the prompts you targeted, position improves. This is the tier that moves first and the one agencies love to report, so keep it honest by demanding the per-engine split.
• Month 4 to 8. The pipeline connection. AI-referred sessions grow enough to read, self-reported attribution starts naming AI on discovery calls, and you can begin tying visibility to actual conversations.
• Month 8 to 12 and beyond. Revenue attribution firms up. You now have enough closed and influenced pipeline to compute the ROI formula with real inputs instead of estimates.

Our own quarter compressed the visibility tier, moving from 0.18% to 10.35% share of answer in about 90 days. Toku's arc ran the other way, deep and durable: 18 months as a growth partner to reach 86% visibility at position 2.4 on its core prompt. Both are real, and they prove the tiers measure different clocks. A fast visibility win is a different thing from a durable revenue engine, and a program judged on visibility alone will always look finished long before the money shows up.

Questions to ask your AEO agency about reporting

Print these. Ask them before you sign, and again every quarter.

1. Do you report all four tiers, and can you show me last month's numbers for a current client (redacted is fine)?
2. Do you break visibility out per engine, or hand me one blended figure?
3. Which metric do you call the headline, visibility or share of voice, and can you state the difference on the spot?
4. How do you separate AI-driven growth from overall market growth?
5. How do you handle the Direct-bucket problem and native-app traffic that carries no referrer?
6. What is your attribution stack, tool by tool, and what does each one miss?
7. When do you expect revenue attribution to produce a defensible ROI number?
8. Can you show the ROI formula you use, with the inputs behind it, rather than only the output percentage?

An agency that answers these cleanly is measuring. One that redirects you back to a single rising line is reporting.

The tools, briefly

You do not need all of them, and no single tool sees everything. Peec is our internal stack for per-engine visibility, citations, and share of answer. Profound covers similar answer-engine metrics with per-platform breakdowns. Ahrefs Brand Radar reports brand mentions across engines and is moving its demand estimate to an AI-adjusted "ask volume," retiring the old keyword-volume basis on August 31, 2026. Semrush scores AI visibility on a 0 to 100 scale against competitors. Every one of them queries the engines from the outside and estimates, which is exactly why your own server logs and CRM are the two highest-fidelity signals you own.

Measurement is not the goal. It is the evidence. If your agency cannot walk the ladder from crawler logs to closed revenue and show you where every number comes from, the ROI they are claiming is a number they made comfortable, not one they earned. Make them earn it. If you want that standard built into a program from day one, that is what our GEO agency work is for.