ACO Measurement: Designing KPIs for AI Shopping Agent Selection Rate and Agent-Assisted Conversion
A five-layer KPI model for ACO: visibility, selection, access, purchase path, and conversion. Covers formulas for agent selection rate and agent-assisted conversion rate, the GA4 ai-assistant medium, signed-request verification, and a measurement tool comparison on one page.
"What should we actually measure for ACO?" Commerce brands run into this question soon after they start on ACO (agentic commerce optimization). A GEO dashboard's mention rate and citation share only tell you whether the agent saw your brand. Whether it picked the product, read the product page, or finished checkout lives in other data. This guide splits ACO KPIs into five layers (visibility, selection, access, purchase path, and conversion) and sets a formula, data source, and collection method for each. Two metrics carry the most weight: agent selection rate and agent-assisted purchase conversion rate.
30-second definitions: the five core ACO KPIs
- Share of prompts is a visibility metric that measures how often a brand appears in AI recommendations across a fixed set of shopping questions (Profitero+, 2026)[3].
- Agent selection rate is a selection metric that measures the share of responses in which the agent names the brand or product as its final recommendation (Enrich Labs, 2026)[1].
- Agent readability rate is an access metric that measures the share of product pages whose HTML, as fetched by the agent, actually contains the price, stock status, and add-to-cart button (Enrich Labs, 2026)[1].
- Purchase path completion rate is a path metric that measures the share of attempts in which the agent gets through every step: offer check, price check, add to cart, and checkout (Enrich Labs, 2026)[1].
- Agent-assisted conversion rate is an outcome metric that measures the share of sessions or order attempts arriving via an AI agent that end in a completed purchase.
The last definition has no citation because the industry has no standard definition for it yet. The formulas in this guide are a proposed design that packages measurement items from public sources into KPI form. Numeric targets are left out on purpose.
ACO KPI design table: formulas, data sources, collection methods
| Layer | KPI | Formula (proposed) | Data source | Collection method |
|---|---|---|---|---|
| 1 Visibility | Share of prompts | Responses featuring the brand / measured responses | AI engine responses | Repeated runs of a fixed prompt set |
| 1 Visibility | AI share of voice (Google) | Your share of voice as calculated by Merchant Center | Merchant Center AI performance insights | Report lookup |
| 2 Selection | Agent selection rate | Responses with a pick / measured responses | AI engine responses | Score each response as picked, mentioned, or absent |
| 2 Selection | Product card rate | Responses showing a product card (image, price) / responses featuring the brand | AI engine responses | Log the display format |
| 3 Access | Agent readability rate | Product pages with price, stock, and cart in the HTML / pages checked | Site HTML | Render checks |
| 3 Access | Verified agent requests | Agent requests that pass signature verification | Server logs, WAF | Web Bot Auth signature verification |
| 4 Path | Purchase path completion rate | Tests reaching the checkout step / path tests run | Path test records | Step checks for offer, price, cart, and checkout |
| 5 Conversion | Agent-assisted conversion rate | Completed AI-assisted purchases / AI-assisted sessions or order attempts | GA4, order system | Medium ai-assistant, sales channel value |
| 5 Conversion | Revenue per AI-assisted visit | AI-assisted revenue / AI-assisted sessions | GA4 | Ecommerce reports |
Each formula turns a measurement item from public sources into a ratio. Enrich Labs lists picked/mentioned/absent scoring, step-by-step purchase path checks, and render checks as ACO measurement items (Enrich Labs, 2026)[1]. Paz.ai's definition of ACO covers logging product cards separately from plain mentions and running the same questions across multiple engines on a fixed schedule (Paz.ai, 2026)[2]. Merchant Center calculates Your share of voice as your brand's AI visibility divided by total visibility, including the competitor set defined by Merchant Center. Merchants cannot change that set directly (Google Merchant Center Help, 2026)[4].
The evidence: why mention rate alone falls short
Start with conversion. Adobe data shows AI traffic to US retail sites grew 393% year over year in Q1 2026. In March 2026, AI traffic converted 42% better than non-AI traffic and generated 37% more revenue per visit. A year earlier, in March 2025, AI traffic converted 38% worse (Adobe via TechCrunch, 2026)[7]. The direction flipped within a single year, so a team that watches only visibility metrics and leaves the conversion layer empty is missing a large share of the results.
The same data exposes a problem at the access layer. In that Adobe analysis, AI could not properly access about 34% of individual product pages (Adobe via TechCrunch, 2026)[7]. An agent can pick your brand and still fail to buy if it cannot read the price and stock status on the product page. That failure never shows up on a mention-rate dashboard, which is why layer 3 gets its own readability metric.
Visibility also tracks poorly with traditional search performance. In an early Profitero+ analysis of 36,000 Amazon products, only 18% of products on page one of standard search also appeared in Rufus results (Profitero+, 2026)[3]. In the same post, Profitero+ wrote that brands now need to measure how often they appear in AI recommendations for common questions and use cases, in other words share of prompts (Profitero+, 2026)[3]. SEO rankings make a weak proxy for layer 1.
Implementing each layer: what to collect and how
Layers 1 and 2: lock down the scoring rules first
Selection rate means nothing if the scoring rules drift. Record each response as exactly one of three values. Picked means the agent presented the brand or product as its recommendation. Mentioned means the name appeared only as a comparison candidate or in an explanation. Absent means it did not appear. Enrich Labs applies this three-way scoring daily per engine and treats a pick as the success condition for ACO (Enrich Labs, 2026)[1]. Selection rate counts only picks in the numerator, while share of prompts counts picks plus mentions. Split this way, GEO and ACO metrics come from the same raw data without overlapping.
Log the display format from the same response too. A product card with an image and price and a passing mention inside a sentence carry different odds of a click and a purchase (Paz.ai, 2026)[2]. Prompt set design and per-engine variance are covered in GEO KPIs and prompt monitoring and multi-engine measurement. For why tools report different numbers, see AI visibility tool measurement discrepancies.
For Google surfaces, check Merchant Center before running your own prompts. The AI performance insights report shows Your share of voice in AI Mode and AI Overviews, the competitor average share, and the number of products showing for top queries (Products showing). It counts only organic AI traffic and leaves out paid ads (Google Merchant Center Help, 2026)[4]. Setup steps are in Merchant Center AI performance insights and UCP.
Layer 3: separate agent visits in your server logs
Readability rate checks two things. The first is whether a product page's initial HTML contains the price, stock status, and add-to-cart button. Values that appear only after JavaScript runs can be missed by an agent. For Product structured data, reuse the checklist in the AEO structured data schema guide.
The second is whether real agent requests are arriving. OpenAI attaches Signature and Signature-Input headers, plus a Signature-Agent header with the value "https://chatgpt.com", to cloud browser requests from ChatGPT agent, and publishes its public keys at chatgpt.com/.well-known/http-message-signatures-directory. It tells merchants to verify the signature using the RFC 9421 method before granting any allowlist exception (OpenAI, 2026)[6]. According to the same document, Cloudflare identifies this traffic with the bot tag chatgpt-agent, Akamai identifies it as ChatGPT Agent, and both verify the signature automatically (OpenAI, 2026)[6]. Counting only requests that pass signature verification gives you a cleaner denominator than counting user-agent strings, which spoofed bots can fake.
Layer 4: test the purchase path step by step
The point of completion rate is to record where each attempt fails, not just the outcome. Break the path into four steps (offer check, price check, add to cart, and checkout) and log where it broke. Enrich Labs runs real agents through these steps one at a time and reports the point where the path breaks (Enrich Labs, 2026)[1]. The same source sorts failures into three types: the agent never found the brand, found it but could not verify it, or got stuck at the purchase step (Enrich Labs, 2026)[1]. Use that classification as is, and you know right away whether to fix layer 1, 3, or 4.
Bot blocking, fraud detection, and CAPTCHA rules often stop agents at checkout. A checklist for the payment flow is in the ChatGPT Wallet agent payments analysis.
Layer 5: combine the GA4 medium with the order channel value
Use GA4 for paths that leave a site session. Under Google's help center definition, when the referrer matches the AI assistant list, the medium is set to ai-assistant and the campaign to (ai-assistant), and the AI Assistant channel groups traffic from sources such as ChatGPT, Gemini, DeepSeek, Copilot, and Grok. The channel excludes Google AI Overviews and AI Mode (Google Analytics Help, 2026)[5]. Take sessions with this medium as the denominator for agent-assisted conversion rate and purchase events in those sessions as the numerator. For channel setup and dark traffic adjustments, follow the GA4 AI Assistant channel guide.
When an agent handles checkout for the user, GA4 may record no session. The order system's sales channel value becomes the source of truth here. ACP is an open standard that OpenAI and Stripe codeveloped and released in September 2025; it lets merchants sell through AI agents while keeping control of their products, branding, and fulfillment (Stripe, 2025)[8]. UCP, which Google released in January 2026, standardizes the full commerce journey, from discovery and evaluation through purchase and order management (Google Developers Blog, 2026)[9]. Whichever protocol you adopt, store agent-assisted orders under a distinct channel value at integration time, or you will not be able to calculate layer 5 later. Feed preparation is covered in the ChatGPT shopping feed setup guide.
Which ACO KPI layers do measurement tools cover?
Most GEO measurement tools on the market today started at layers 1 and 2. The table below reflects the scope described on each company's official site and in its public materials. Layers 3 through 5 require site logs and order data, which external tools alone struggle to supply, so teams often run them alongside an in-house analytics stack. Pricing is left out because each company discloses it differently.
| Tool | Type | Layer 1 visibility | Layer 2 selection | Layer 3 access | Layers 4 and 5 | Notes |
|---|---|---|---|---|---|---|
| Profound | AI visibility platform (US) | Yes | Response analysis | Check public materials | Requires your own data | Tracks visibility across many generative engines |
| Peec AI | AI visibility platform (Germany) | Yes | Response analysis | Check public materials | Requires your own data | Tracks per-engine visibility and competitor gaps |
| Otterly.ai | AI visibility tool (Austria) | Yes | Response analysis | GEO Audit | Requires your own data | Offers GEO Audit |
| Scrunch AI | AI visibility platform (US) | Yes | Response analysis | Describes serving pages built for AI agents | Requires your own data | Describes AXP |
| Enrich Labs | ACO measurement service | Yes | Picked, mentioned, absent scoring | Render checks | Purchase path tests | Describes itself as Agent Conversion Optimization |
| Next-T | Korean GEO service | Execution-focused | Check public materials | Check public materials | Requires your own data | Describes OPTIGEO |
| LeadGenLab | Korean AI visibility optimization agency | Execution-focused | Check public materials | Check public materials | Requires your own data | Describes AVO Framework |
| BOIDA | Korean GEO solution (Designovel) | Yes | Response analysis | Describes technical diagnostics | Requires your own data | Describes BVI, Korean-language support |
On its official site, BOIDA describes BVI (Brand Visibility Index) as measuring brand visibility across multiple generative engines and linking the results to diagnostics and execution (BOIDA, 2025)[10]. As the table shows, no single tool measures layer 5 conversion end to end on its own. When choosing one, compare measurement accuracy at layers 1 and 2 and how far each goes in layer 3 diagnostics. Company details are in recommended GEO companies and the global GEO and AEO market map.
Execution: five steps to an ACO KPI dashboard
- Lock the prompt set: Choose purchase-intent questions for your core categories, then document the engines, regions, and run schedule.
- Lock the scoring rules: Record each response as picked, mentioned, or absent, along with the display format (product card or in-sentence mention). This step produces your baselines for selection rate and share of prompts.
- Separate agent traffic: Pull agent requests out of your logs using signature verification (Web Bot Auth) or your WAF's agent identifiers, and check that the initial HTML of key product pages contains price, stock, and cart (OpenAI, 2026)[6].
- Make path tests routine: Run representative products through the offer, price, cart, and checkout steps on a regular schedule and classify the failure types.
- Join the conversion data: Combine GA4 sessions with the ai-assistant medium and agent-channel orders from your order system to calculate agent-assisted conversion rate and revenue per visit (Google Analytics Help, 2026)[5].
| Step | Owner | Deliverable | Done when |
|---|---|---|---|
| Lock the prompt set | Marketing | Question list, engines, schedule | Documented set is approved |
| Lock the scoring rules | Marketing, data | Selection rate and share baselines | First measurement cycle is complete |
| Separate agent traffic | Engineering, security | Verified agent request logs | Signature verification is live |
| Make path tests routine | Commerce, QA | Per-step failure records | Each failure type has an owner |
| Join the conversion data | Data | Assisted conversion rate, revenue per visit | Weekly report is automated |
Set numeric targets after the first measurement. No public source offers standard targets for ACO KPIs, and Profitero+ likewise defines the metrics and the digital shelf basics without setting numeric goals (Profitero+, 2026)[3]. For how this ties into your broader commerce strategy, see agentic commerce GEO strategy.
Frequently asked questions
How is agent selection rate different from the GEO mention rate?
Mention rate counts whether the name appeared. Selection rate counts whether it was named as the final recommendation. Each response is scored as picked, mentioned, or absent (Enrich Labs, 2026)[1].
Does GA4 capture every agent-assisted purchase?
No. GA4 assigns the ai-assistant medium based on the referrer, so when an agent handles checkout for the user, no session may be recorded. Read the order system's channel value and your server logs alongside it (Google Analytics Help, 2026)[5].
Where does traffic from Google AI Mode get counted?
Not in the GA4 AI Assistant channel. Track AI Mode visibility separately through the share metrics in Merchant Center AI performance insights (Google Analytics Help, 2026; Google Merchant Center Help, 2026)[5][4].
How often should selection rate be measured?
Run the same prompt set on a fixed schedule and watch the trend. The change across repeated runs under the same conditions is a better basis for decisions than any single reading (Paz.ai, 2026)[2].
Is there an industry-standard target for ACO KPIs?
No validated standard target exists yet (Profitero+, 2026)[3]. Treat your first reading as the baseline and set targets as improvement over it.
Takeaways
Mention rate alone cannot measure ACO performance. Visibility and selection data come from AI responses, access and purchase path data from site HTML and server logs, and conversion data from GA4 and the order system. The two metrics to stand up first are agent selection rate and agent-assisted conversion rate. You can measure selection rate as soon as the scoring rules are fixed, and assisted conversion rate becomes calculable once the GA4 ai-assistant medium and the order channel value are in place. When the gap between the two widens, the problem sits in layers 3 and 4: product page readability and the checkout path. If you need the concepts first, start with ACO vs. GEO vs. AEO.
Related companies
- 넥스트티 (Next-T, OPTIGEO)SEO, GEO, AEO 컨설팅, 자동화
- 리드젠랩 (LeadGenLab)AI 가시성 최적화 에이전시
- 보이다 (BOIDA)생성형 검색 최적화(GEO) 솔루션, AI 가시성 측정
- Otterly.aiAI 가시성 모니터링 툴
- Peec AIAI 가시성 모니터링 플랫폼
- ProfoundAI 가시성 모니터링 플랫폼
- Scrunch AIAI 가시성 모니터링 플랫폼
Frequently asked questions
- Mention rate counts whether the brand name shows up in a response. Selection rate counts whether the agent named the brand as its final recommendation. Enrich Labs scores each response as picked, mentioned, or absent and treats a pick as the success condition for ACO (Enrich Labs, 2026).
- No. GA4 classifies the medium as ai-assistant when the referrer matches its list of AI assistants, but if an agent completes checkout on the user's behalf, no site session may be recorded at all. Read the sales channel value in your order system and your server logs alongside GA4 (Google Analytics Help, 2026).
- Not in the GA4 AI Assistant channel. Google's help center states that this channel excludes Google AI Overviews and AI Mode. Track brand visibility in AI Mode separately through the share metrics in Merchant Center AI performance insights (Google Analytics Help, 2026; Google Merchant Center Help, 2026).
- Run the same prompt set across multiple engines, such as ChatGPT, Google AI Mode, and Perplexity, on a fixed schedule, and judge by the trend across repeated runs rather than any single reading (Paz.ai, 2026).
- Not yet, at least none that has been validated. Public sources define the metrics and the measurement methods but do not set numeric targets (Profitero+, 2026). The practical approach is to treat your first reading as the baseline and set targets as improvement over it.
Q.How is agent selection rate different from the GEO mention rate?
Q.Does GA4 capture every agent-assisted purchase?
Q.Where does traffic from Google AI Mode get counted?
Q.How often should selection rate be measured?
Q.Is there an industry-standard target for ACO KPIs?
Sources
- [1] ↑Agent Conversion Optimization (ACO): Get Recommended and Bought by AI Shopping Agents — Enrich Labs
- [2] ↑Agentic Commerce Optimization (ACO) — Paz.ai
- [3] ↑Commerce 101: How AEO, GEO and the Digital Shelf are Shaping the Next Wave of Commerce — Profitero+
- [4] ↑About AI performance insights — Google Merchant Center Help
- [5] ↑Default channel group — Google Analytics Help
- [6] ↑Help ensure cloud browser access to your site — OpenAI
- [7] ↑AI traffic to US retailers rose 393% in Q1, and it's boosting their revenue too — TechCrunch
- [8] ↑Stripe powers Instant Checkout in ChatGPT and releases Agentic Commerce Protocol codeveloped with OpenAI — Stripe
- [9] ↑Under the Hood: Universal Commerce Protocol (UCP) — Google Developers Blog
- [10] ↑BOIDA BVI official site — Designovel
Related documents
- ACO vs. GEO vs. AEO: What Agentic Commerce Optimization Adds on Top (2026)How ACO (agentic commerce optimization) differs from GEO and AEO in goal, optimization target, measurement endpoint, and technical requirements. Covers the ACP and UCP protocols, Adobe traffic data, a rollout sequence, and a solution comparison table on one page.
- GEO KPI Performance Measurement: Citation Rate, AI SoV, and Prompt MonitoringA step-by-step guide to the core GEO KPIs: citation rate, AI Share of Voice, and prompt monitoring. Covers the arXiv-grounded statistical framework, brand-scale benchmarks, and a side-by-side comparison of domestic and global measurement tools for quantifying AI search visibility.
- GA4 AI Assistant Channel: Measuring ChatGPT, Gemini, and Claude TrafficGA4's AI Assistant channel: added to Default Channel Groups in May 2026, automatically classifies AI referral traffic, yet a structural blind spot remains: sessions that arrive without a referrer header. Here's what the channel captures, what it misses, and how to cover the gap.
- Google Merchant Center AI Performance Insights: AI Share of Voice and the 6 Conversational AttributesWhat the four AI performance insights metrics actually count, the filters and edge cases that break interpretation, how to submit the six conversational attributes, what changed in the UCP integration hub, and what Korean accounts can use instead. Built from Google's Help documentation.
- ChatGPT Wallet Launch: What Agent Payments Mean for Brands and Commerce GEOA primary-source breakdown of ChatGPT Wallet, which OpenAI unveiled in November 2026: how it works (payment tokens, per-purchase approval), who can use it, and what merchants need, plus how agent payments change brand visibility and the commerce GEO checklist.
- ChatGPT Shopping Feed Setup: From Merchant Portal to AI Product VisibilityA hands-on guide for sellers who want their products appearing in ChatGPT Shopping results. Covers chatgpt.com/merchants access, required feed fields, common rejection causes, and how to measure AI visibility after approval.
- Agentic Commerce GEO: How to Enter the Shopping Agent's Selection PoolAI shopping agents now browse, compare, and complete purchases without explicit user commands. Getting into their consideration set requires structured product data and protocol connectivity. This guide covers ACP, UCP, and MCP protocols, Naver's AI shopping agent, and the GEO execution steps brands should act on now.
- Multi-Engine Measurement: How to Measure Visibility Across ChatGPT, Gemini, Perplexity, and ClaudeWhy every engine answers differently, the trap of single-engine measurement, and a multi-engine GEO methodology for measuring AI visibility through prompt sets, repetition, and share of voice.
- Why AI Visibility Tools Report Different Scores for the Same Brand: Causes and a Verification ProcedureWhy the same brand scores differently in every AI visibility tool, split into four layers (engine variance, collection design, metric definition, aggregation), with a side-by-side table of the published formulas from Profound, Peec AI, and Ahrefs, sample-size confidence interval thresholds, and a six-step verification procedure.
- Structured Data and Schema Guide for AEOStructured data (JSON-LD) from schema.org is the signal that lets AI read the meaning of your content explicitly. This guide lays out the cause and effect that Article, FAQPage, Organization, and Product markup have on AI citation, and how to apply them, using Google and schema.org sources with JSON-LD examples.
- Global GEO/AEO Player Landscape 2026, Monitoring Tools, Agencies, and PlatformsA 2026 landscape that sorts GEO/AEO players into monitoring tools, specialist solutions and agencies, enterprise platforms, and regional players. We compare the leading vendor in each category, founding, headquarters, tracked engines, pricing, and differentiation, against primary sources.
- Best GEO/AEO Companies: Domestic & Global Agencies and Solutions Comparison Guide 2026A comprehensive answer to 'which GEO companies are worth recommending.' Compares domestic Korean (Intermajor, ZESTCOMPANY, Narr/Answer, BizSpring, etc.) and global monitoring tools, diagnosis solutions, and agencies by founding, headquarters, and differentiators.