---
title: "By 2028, AI Visibility Reports Without Denominators Will Be Worthless"
description: "Three checkable predictions for AI visibility reporting by 2028: fixed denominators, separate recommendation and citation rates, and honest attribution."
url: "https://groas.com/post/by-2028-an-ai-visibility-report-without-a-denominator-will"
image: "https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/0fe3e7fc-7c5d-457d-a7ae-6bd38c0c4012.png"
published: "2026-11-07T06:09:52.913Z"
modified: "2026-11-10T21:20:11.047Z"
---

November 7, 2026 · 10 min read

# By 2028, AI Visibility Reports Without Denominators Will Be Worthless

[](https://groas.com/author/alexander-perelman)

[Alexander Perelman](https://groas.com/author/alexander-perelman)Personally Tested

Head Of Product @ groas

[LinkedIn](https://www.linkedin.com/in/alexander-433793253/)

Updated November 10, 2026

![A gleaming gold '42' sits atop a floating fraction bar, but the plinth below where the denominator belongs is empty, a count with nothing to measure it against.](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/0fe3e7fc-7c5d-457d-a7ae-6bd38c0c4012.png)

In this article

1. [Prediction 1: buyers will demand fixed prompt panels and repeat observations](#prediction-1-buyers-will-demand-fixed-prompt-panels-and-repeat-observations)
2. [Prediction 2: recommendation and citation rates will become separate buying criteria](#prediction-2-recommendation-and-citation-rates-will-become-separate-buying-criteria)
3. [Prediction 3: credible commercial reports will disclose attribution uncertainty](#prediction-3-credible-commercial-reports-will-disclose-attribution-uncertainty)
4. [What I would put in a managed-service contract now](#what-i-would-put-in-a-managed-service-contract-now)

By December 31, 2028, an AI visibility report that gives you a count without a denominator will be worthless. **If your vendor reports 42 mentions but cannot tell you “42 out of how many eligible runs,” you have no evidence of progress.**

I used to sell that kind of counting myself in paid search. More impressions felt like more visibility until you divided by eligible impressions and found share flat. AI answers have the same flaw, only worse: the same buyer prompt asked twice can name different brands and flip the recommendation because of sampling and retrieval variance [across runs](https://www.menra.ai/glossary/prompt-sampling).

**Free tool:** the [AI Visibility Checker](https://groas.com/free-tools/ai-visibility-checker) asks ChatGPT, Gemini, Claude, Perplexity and Grok the question your customers ask and shows whether they recommend you.

One screenshot proves even less. In one widely cited test, ChatGPT and Google AI returned the same brand list in fewer than 1 in 100 repeat runs of the same prompt, with [identical order closer to 1 in 1,000](https://geocitely.com/blog/ai-mentions-vs-citations). A screenshot records an answer. It does not establish a pattern.

My forecast for the next two years is simple: sophisticated buyers will stop accepting raw mentions as evidence of progress. They will demand repeatable recommendation rates, separate source-citation rates and attributable enquiries, with a defined denominator for each.

The problem starts with who controls the test. If I get paid on mentions, I can add ten soft prompts where you show up anyway, run each once and send you a bigger number. Your buyer visibility has not changed. The prompt panel has. Credible AI share of voice therefore needs to be [a rate over a declared prompt set, not a platform number](https://www.aicitekit.com/blog/ai-search-share-of-voice-denominator/), with the numerator, denominator, surfaces, dates and exclusion rules published.

![Illustrated scoreboard showing 42 mentions without a denominator, circled as meaningless](https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/ee7b0266-2074-4ef1-9ece-64f461375cf0.png)

A mention, a citation and a recommendation answer different questions. On the same illustrative set of 40 eligible runs, a report might show:

- **Named:** 18 of 40 runs, or 45 percent.
- **Recommended:** 9 of 40 runs, or 22.5 percent.
- **Cited:** 7 of 40 runs, or 17.5 percent.

Those are not interchangeable wins. A brand can have a high mention rate and a low citation rate, or get cited in an answer explaining a competitor. [Keep the measures separate](https://www.aicitekit.com/blog/ai-search-share-of-voice-denominator/). Otherwise, the vendor gets to decide what _visibility_ means after seeing the results.

## Prediction 1: buyers will demand fixed prompt panels and repeat observations

By 2028, I expect a serious procurement document to name the engines, buyer situations, sampling frequency and denominator before it names the price.

The mechanism is simple. A Monday check says visibility is up; Friday says it is down. If both are one-shot observations, neither tells you whether anything changed. Once buyers have paid for enough of that noise, they will stop buying screenshots and start demanding distributions.

Say you spend $20k a month on search. A measurement specification could look like this:

1. Lock and version a panel of 40 buyer prompts.
2. Run it weekly on ChatGPT, Gemini and Perplexity.
3. Repeat each prompt three to five times for stable questions and ten or more for contested ones.
4. Report outcomes as counts and rates, such as 7 of 10 eligible runs, with the interval shown.
5. Disclose changes to the panel rather than quietly folding them into the trend.

That design already exists in [prompt sampling guidance](https://www.menra.ai/glossary/prompt-sampling). Repeated observations address the sampling, retrieval and context variance that can change an answer from one run to the next. The point is not to make the report look scientific. It is to avoid paying someone to explain random movement.

**Every reported rate needs an inspectable denominator.** A minimum report row should include:

- The metric, numerator and denominator.
- The prompt cohort and engine or surface.
- The market and reporting window.
- The eligibility and exclusion rules.

“9 recommendations in 40 eligible runs” is a start. It becomes useful when the report also publishes [what made those runs eligible and what was excluded](https://www.aicitekit.com/blog/ai-search-share-of-voice-denominator/).

This is not a specification buyers have to invent from scratch. Managed programs covering ChatGPT, Gemini, Perplexity, Copilot and AI Overviews already sell [dated baselines on 30 to 100 buyer prompts, with tracking tied to a fixed set](https://www.ogtool.com/blog/generative-engine-optimization-services-what-you-buy-what-you-measure-and-what-actually-gets-cited-in-2026), typically from $3,000 to $25,000 a month. The honest ones will lock the panel and show repeats. The rest will add prompts until the count looks good.

**What would prove me wrong:** by late 2028, most winning proposals still report raw mentions without a prompt list, run count or exclusions, and buyers renew them anyway. If counting keeps winning the contract, my prediction fails.

Ask for the measurement row before the sales demo.

## Prediction 2: recommendation and citation rates will become separate buying criteria

Being used as a source does not mean making the shortlist. A citation points to a source. A recommendation selects an option. I expect buyers to stop letting vendors combine those outcomes into one flattering score.

The distinction matters because sourcing and selection can diverge. In one 30-prompt test on the OpenAI API, Semrush was named in 10 prompts, but its own site was cited in only 6 of those 10. Trakkr was [cited once without ever being named](https://geocitely.com/blog/ai-mentions-vs-citations).

The split also appears in larger datasets. One June 2026 dataset found [61.7 percent of brand appearances were citations where the brand was never named](https://geocitely.com/blog/ai-mentions-vs-citations). A study of 12,000 answers found that [only 23.1 percent of brand mentions came with an own-site citation](https://geocitely.com/blog/ai-mentions-vs-citations).

That is the evidence today. The buying consequence comes next: a report that blends these outcomes cannot tell me whether I have a content problem or a preference problem. Am I supplying material the answer uses without becoming an option it recommends? Am I getting named without my own content supporting the answer? Those call for different work.

**I want recommendation rate and citation rate scored separately, over the same eligible runs.** Keep the named-brand rate alongside them. Do not let one stand in for another merely because it moved in the right direction.

The next requirement is less glamorous: a change must survive repetition before it earns the word _improvement_. That means observing the same locked prompt, engine and buyer situation again, not finding a fresh screenshot.

Sampling guidance already calls for [three to five runs for stable prompts and ten or more for contested ones](https://www.menra.ai/glossary/prompt-sampling), reported as distributions such as 7 of 10 runs. I would add a contract rule: no claim of improvement until the lift holds across two consecutive sampling windows.

That rules out the oldest trick in this category: showing the Tuesday run where you ranked first and ignoring the four runs where you did not. It also connects measurement to the work. A citation gap and a recommendation gap should not disappear into the same task labelled “improve AI visibility.” That is a heading, not an explanation.

**What would prove me wrong:** by 2028, buyers still renew vendors on a single blended visibility score, with no split between named, cited and recommended and no repeat runs disclosed. If buyers keep paying without the distinction, my prediction fails.

Put the split and the repeatability rule in the contract, not a footnote in the report.

## Prediction 3: credible commercial reports will disclose attribution uncertainty

Here is where most AI visibility decks fall apart: Direct traffic rises, AI mentions rise, and the deck declares that the second caused the first. I have run that play in paid search, with branded queries taking credit for demand they did not create.

With AI answers, the attribution gap is wider. An estimated [70 percent of traffic influenced by AI assistants never shows up as AI in analytics](https://www.influencers-time.com/dark-ai-traffic-hides-in-direct-ga4-still-misses-it/). It can land in Direct or branded organic, or produce no click at all. GA4 added a native AI Assistant channel in May 2026, but it only classifies visits where the referrer survives. Mobile webviews, copy-paste answers and paid-tier sessions can still fall into Direct.

There is commercial value worth measuring here. Adobe looked at more than 1 trillion visits and found [AI referral traffic up 393 percent year over year in early 2026](https://www.influencers-time.com/dark-ai-traffic-hides-in-direct-ga4-still-misses-it/), converting 42 percent better, with revenue per visit 37 percent above non-AI traffic. That tells me observed AI-assisted visitors can be valuable. It does not identify which untracked visit came from which answer.

The mechanism behind my prediction is that gap: buyers will want commercial evidence, but the available tracking will not justify assigning every enquiry to a specific AI interaction. Credible services will have to show their limits rather than fill them with optimistic arithmetic.

**An unknown is a reporting category, not permission to claim the revenue.** By 2028, I expect a credible report to distinguish:

- **Observed AI referrals:** visits identifiable as coming from an assistant.
- **Enquiries with supporting evidence:** CRM enquiries where buyer language matches the prompt panel, without pretending that match proves the whole journey.
- **Unknown influence:** activity the available evidence cannot assign.

Keep Google Search performance beside those lines. It is a check against a vendor hiding a Google loss inside an AI win.

Google’s generative AI performance report in Search Console [launched June 3, 2026 and reached all properties August 31, 2026](https://www.cognizo.ai/blog/track-google-ai-overviews-traffic-and-click-loss-gsc-guide). It reports impressions by page, country, device and date, but not clicks, CTR, position or queries. It also lacks a split between AI Overviews and AI Mode and has no API access.

Clicks from AI Overviews sit inside overall Web totals without a key to join them to those impressions. An [AI-specific CTR therefore cannot be calculated from Search Console alone](https://www.cognizo.ai/blog/track-google-ai-overviews-traffic-and-click-loss-gsc-guide). A vendor presenting exact AI revenue to the dollar from that setup is filling the gap with assumptions. Those assumptions belong in the report, not behind the curtain.

**What would prove me wrong:** by 2028, Google and the assistant engines expose click-level joins that connect a specific answer to a specific enquiry without modelling. If full attribution becomes directly checkable, my uncertainty rule fails.

Until then, demand evidence of contribution without accepting invented certainty.

## What I would put in a managed-service contract now

If I were buying ongoing management across ChatGPT, Gemini and Perplexity this quarter, I would require three things before signing:

1. **A stable measurement specification.** The versioned prompt list, engines, buyer situations, repeats per prompt, reporting cadence and denominator for every rate.
2. **An execution log.** Each change tied to the recommendation or citation gap it was meant to close.
3. **Explicit limits.** Out-of-scope prompts, eligible-run rules and the activity the report will label unknown rather than assign to AI.

The second requirement matters as much as the first. A monitoring tool can tell me that I appeared in 9 of 40 eligible runs. It does not do the work intended to change the next 40. With tracking alone, I pay a subscription and keep the job. With management, I am paying for catalogue fixes, quotable content units and third-party corroboration that give models something to cite.

That is the test I apply to [groas](https://groas.com/). It is sold as done-for-you management for paid ads, SEO and AI search, not another dashboard I have to interpret. Its structure pairs continuous execution with a named human who owns the outcome, with every action logged alongside its reasoning. That is a stronger operating model than a monthly account review followed by a percentage-of-spend invoice.

The reporting bar should match it. I want separate recommendation and citation rates over the same fixed denominator, movement claimed only after two sampling windows, and Google Search performance alongside the AI lines. Continuous work deserves a record that makes the work inspectable.

**Before the next renewal, I would run this test for eight weeks:** freeze 30 to 50 prompts buyers actually ask, lock the engines and weekly run count, and require separate named, cited and recommended rates. Put the repeatability rule in writing. Keep enquiries and observed referrals beside the rates, with the rest in an explicit unknown bucket.

Vendors who live on inflated counts will fight the specification. The ones doing the work will sign it, because it lets them demonstrate progress and put visibility beside the outcomes that matter: CPA, pipeline and revenue, not screenshots.

Check me on December 31, 2028. If denominators are still missing and buyers still renew, I was wrong and I will say so.

## Frequently Asked Questions

### Why is a raw AI mention count without a denominator considered insufficient for tracking visibility?

A raw count is insufficient because it does not show visibility relative to the total possible opportunities. To prove progress, a vendor must show the count against the total number of eligible runs, which provides necessary evidence of share.

### How can businesses stabilize AI visibility reporting when the answers change between different runs?

Visibility must be measured using fixed prompt panels and repeatable observations. Best practices recommend running prompts three to five times for stable questions and ten or more times for contested questions to minimize variance.

### Why must AI report categories like mentions, citations, and recommendations be tracked separately?

These three metrics answer different questions, such as whether a brand is mentioned, cited as a source, or selected as an option. Combining them prevents a vendor from hiding a content problem behind a visibility rate that is only moving in the right direction.

### What should be disclosed when tracking the influence of AI assistants on consumer traffic?

Credible reports must distinguish between observed AI referrals, inquiries with supporting evidence, and unknown influence. This acknowledges the gap where untracked activity, such as copy-pasting answers, cannot be assigned to a specific AI interaction.

### What key elements should be required in a managed service contract for AI visibility?

A strong contract requires a stable measurement specification, an execution log detailing all changes, and explicit limits. This ensures that visibility efforts are backed by a predictable methodology and fully traceable actions.

## Run Your Marketing On groas

Plug ChatGPT, Claude or any AI agent into groas and let it do the work of a whole marketing team: free to ask, $49 a month flat with every feature included, and autopilot whenever you’re ready.

[Start Free](https://groas.com/pricing)

[Or have groas run it for you](https://groas.com/done-for-you)

## Structured data

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://graph.groas.com/entity/groas","name":"groas","url":"https://groas.com/","description":"groas is the marketing infrastructure for your own AI: plug Claude, ChatGPT or any AI agent into groas and run all your marketing from it, for a tiny fraction of the usual cost, with the groas app there whenever you want to see more. It is the layer ChatGPT, Claude and any AI assistant run a business’s online marketing on. Through groas an assistant can do anything you could do natively in Google Ads, Bing Ads, Meta Ads and ChatGPT ads, on your website and in AI search, and, where the owner switches it on, can run connected campaigns by itself, so AI agents do the work a whole marketing team used to (free to read, $49 a month to make changes, delivered as an MCP server). Separately, its done-for-you service has senior people and proprietary quantitative systems run paid ads, SEO and AI search for businesses, and white label for agencies, from $5,000 a month.","logo":{"@type":"ImageObject","url":"https://groas.com/icon-512.png","width":512,"height":512},"sameAs":["https://www.linkedin.com/company/groas/"]},{"@type":"WebSite","@id":"https://groas.com/#website","name":"groas","url":"https://groas.com/","inLanguage":"en-US","publisher":{"@id":"https://graph.groas.com/entity/groas"}},{"@type":"WebPage","@id":"https://groas.com/post/by-2028-an-ai-visibility-report-without-a-denominator-will#webpage","url":"https://groas.com/post/by-2028-an-ai-visibility-report-without-a-denominator-will","name":"By 2028, AI Visibility Reports Without Denominators Will Be Worthless","description":"Three checkable predictions for AI visibility reporting by 2028: fixed denominators, separate recommendation and citation rates, and honest attribution.","inLanguage":"en-US","isPartOf":{"@id":"https://groas.com/#website"},"about":{"@id":"https://graph.groas.com/entity/groas"},"primaryImageOfPage":{"@type":"ImageObject","url":"https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/0fe3e7fc-7c5d-457d-a7ae-6bd38c0c4012.png"},"breadcrumb":{"@id":"https://groas.com/post/by-2028-an-ai-visibility-report-without-a-denominator-will#breadcrumb"},"dateModified":"2026-11-10T21:20:11.047Z","mainEntity":{"@id":"https://groas.com/post/by-2028-an-ai-visibility-report-without-a-denominator-will#article"}},{"@type":"BreadcrumbList","@id":"https://groas.com/post/by-2028-an-ai-visibility-report-without-a-denominator-will#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://groas.com/"},{"@type":"ListItem","position":2,"name":"Blog","item":"https://groas.com/blog"},{"@type":"ListItem","position":3,"name":"By 2028, AI Visibility Reports Without Denominators Will Be Worthless","item":"https://groas.com/post/by-2028-an-ai-visibility-report-without-a-denominator-will"}]},{"@type":"BlogPosting","@id":"https://groas.com/post/by-2028-an-ai-visibility-report-without-a-denominator-will#article","url":"https://groas.com/post/by-2028-an-ai-visibility-report-without-a-denominator-will","mainEntityOfPage":{"@id":"https://groas.com/post/by-2028-an-ai-visibility-report-without-a-denominator-will#webpage"},"isPartOf":{"@id":"https://groas.com/blog#blog"},"headline":"By 2028, AI Visibility Reports Without Denominators Will Be Worthless","description":"Three checkable predictions for AI visibility reporting by 2028: fixed denominators, separate recommendation and citation rates, and honest attribution.","datePublished":"2026-11-07T06:09:52.913Z","dateModified":"2026-11-10T21:20:11.047Z","image":{"@type":"ImageObject","url":"https://pub-87da24ecbbfc4c3bad6875f3aa013712.r2.dev/generated-images/0fe3e7fc-7c5d-457d-a7ae-6bd38c0c4012.png"},"author":{"@id":"https://groas.com/author/alexander-perelman#person"},"publisher":{"@id":"https://graph.groas.com/entity/groas"},"keywords":"visibility, without, 2028, runs, report, prompt, recommendation, citation","wordCount":2036,"inLanguage":"en-US"},{"@type":"Person","@id":"https://groas.com/author/alexander-perelman#person","name":"Alexander Perelman","jobTitle":"Head Of Product @ groas","description":"My background includes Goldman Sachs and Stanford Computer Science.","image":"https://groas.com/media/blog/4f17cac3a81acc224cc1b2cfab8f94bbc295a3089d13e8421125771ef8d4064f.jpg","worksFor":{"@id":"https://graph.groas.com/entity/groas"},"sameAs":["https://www.linkedin.com/in/alexander-433793253/"]},{"@type":"FAQPage","@id":"https://groas.com/post/by-2028-an-ai-visibility-report-without-a-denominator-will#faq","mainEntity":[{"@type":"Question","name":"Why is a raw AI mention count without a denominator considered insufficient for tracking visibility?","acceptedAnswer":{"@type":"Answer","text":"A raw count is insufficient because it does not show visibility relative to the total possible opportunities. To prove progress, a vendor must show the count against the total number of eligible runs, which provides necessary evidence of share."}},{"@type":"Question","name":"How can businesses stabilize AI visibility reporting when the answers change between different runs?","acceptedAnswer":{"@type":"Answer","text":"Visibility must be measured using fixed prompt panels and repeatable observations. Best practices recommend running prompts three to five times for stable questions and ten or more times for contested questions to minimize variance."}},{"@type":"Question","name":"Why must AI report categories like mentions, citations, and recommendations be tracked separately?","acceptedAnswer":{"@type":"Answer","text":"These three metrics answer different questions, such as whether a brand is mentioned, cited as a source, or selected as an option. Combining them prevents a vendor from hiding a content problem behind a visibility rate that is only moving in the right direction."}},{"@type":"Question","name":"What should be disclosed when tracking the influence of AI assistants on consumer traffic?","acceptedAnswer":{"@type":"Answer","text":"Credible reports must distinguish between observed AI referrals, inquiries with supporting evidence, and unknown influence. This acknowledges the gap where untracked activity, such as copy-pasting answers, cannot be assigned to a specific AI interaction."}},{"@type":"Question","name":"What key elements should be required in a managed service contract for AI visibility?","acceptedAnswer":{"@type":"Answer","text":"A strong contract requires a stable measurement specification, an execution log detailing all changes, and explicit limits. This ensures that visibility efforts are backed by a predictable methodology and fully traceable actions."}}]}]}
```
