Case study: Gumlet turned ChatGPT mentions into 20% of inbound revenue. Read it →
Recommendation share sits closer to pipeline than citation share. DerivateX measures both, because a recommendation puts a vendor on a buyer’s shortlist while a citation only proves a page was used as evidence. Across a 150-prompt frame, DerivateX treats citation share as the leading indicator and recommendation share as the revenue-proximate one.
Recommendation share is the percentage of sampled buyer prompts in which an engine names your brand as a vendor worth considering. Citation share is the percentage of source references in those same answers that point to a page you control or influence. One measures whether you made the shortlist. The other measures whether your evidence made the answer.
The gap matters because the two can move in opposite directions. A language model can recommend your product with confidence and cite three review sites, one Reddit thread and a competitor’s comparison page as its sources. Your recommendation share is high and your owned citation share is zero. The reverse also happens: your documentation gets cited constantly for a technical definition, and the engine still recommends someone else when the prompt turns into “which tool should I buy.”
DerivateX keeps them as two separate lines in every client report for that reason. A single blended “AI visibility” number hides the diagnosis. Flat recommendation share with climbing citation share means the corroboration is landing but the vendor framing is not. The opposite pattern means third parties are doing the work and your own pages are not yet part of the evidence base.
| Dimension | Recommendation share | Citation share |
|---|---|---|
| What the numerator counts | Runs where the brand appears in a vendor slot | Source references resolving to your domain, or to pages that name you |
| What it answers | Are we on the shortlist | Is our evidence in the answer |
| Ceiling | Capped by slot count, with 3 to 4 brands recommended per AI category query | Capped by how many sources the engine surfaces per answer |
| Speed of movement | Slower, follows corroboration | Faster, responds to new indexable evidence first |
| Portability across engines | Higher, the vendor set converges | Lower, only 11% of domains are cited by both ChatGPT and Perplexity |
| Distance to pipeline | Closer | Further, but more actionable |
| How DerivateX reports it | Weighted and unweighted, with a 95% interval printed alongside | Split into owned and influenced, at both prompt level and link level |
Recommendation share is closer, and DerivateX will take that position plainly. A recommendation is what enters a buyer’s consideration set. It survives the conversation even when nobody clicks, which is most of the time. Citation share is a supply-side measure of your evidence footprint, and evidence can be used to recommend someone else.
The counterintuitive part is that citation share is still the metric you work on first, because it is the one you can move directly. You cannot write a page that makes an engine recommend you. You can publish, place and correct the evidence that engines read when they assemble a recommendation. That is why the sequence in DerivateX engagements runs citation share, then recommendation share, then arrival, then pipeline, with the lag between each step measured rather than assumed.
Clicks from AI surfaces are scarce but qualified: AI-sourced visitors convert at 4.4x the rate of other sources, and 28% of ChatGPT-cited pages have zero organic Google visibility, so the traffic you do receive is not a rerun of your existing search demand. A recommendation that never produces a click can still surface as a branded search days later, which is a hypothesis to test in your own data rather than a fact to assert. Our AI search funnel measurement model covers that lag structure in more depth, so this article stays on the two share metrics themselves.
You calculate them by fixing four things in writing and never changing them mid-quarter: the prompt frame, the run definition, the numerator rule and the denominator. DerivateX writes these into the measurement doc at kickoff so that any number in a board deck can be recomputed by someone else from the raw logs.
Build a fixed set of buyer prompts, not keywords. For software companies at $5M to $50M ARR, DerivateX typically uses 120 to 200 prompts split across four intents: category selection (“best X for Y”), comparison (“X vs Y”), problem-first (“how do I solve Z”), and qualification (“is X good for a 40-person team”). Lock the wording. Prompt phrasing changes outcomes on its own, so a frame that drifts produces a trend line that means nothing.
A run is one prompt sent to one engine on one date. Total runs equal prompts multiplied by engines multiplied by repeats. Repeats matter because answers are non-deterministic. DerivateX runs each prompt twice per engine per cycle and averages, which is the minimum that produces a usable interval.
Recommendation share uses total runs. Owned citation share can use total runs (prompt-level) or total citations returned (link-level). Pick one and label it. DerivateX reports both, because prompt-level reads high and link-level reads low, and a leadership team that sees one number in Q1 and the other in Q2 will assume something broke.
Unweighted share treats first mention and last mention as equal. They are not. Apply a position weight of 1.0 for the first vendor named, 0.7 for the second, and 0.5 for the third and beyond, then report weighted and unweighted side by side. Those weights are a convention DerivateX chose for stability, not a measured constant, and we label them as such in every report.
Large enough that the move you want to report is bigger than the sampling error. DerivateX computes a 95% interval on every share figure using the normal-approximation interval for a binomial proportion, the standard formula set out in the NIST/SEMATECH e-Handbook of Statistical Methods (2012 revision). Every margin quoted below is DerivateX’s own calculation on that formula, not a published industry benchmark.
On that formula, 150 unique prompts run once puts a margin of roughly plus or minus 6.8 points around a 24% recommendation share, so a jump from 24% to 29% is not reportable. Increase the run count to 1,200 (150 prompts, four engines, two repeats) and the same figure carries a margin of about plus or minus 2.4 points. Getting inside plus or minus 3 points at that share level takes roughly 780 runs, which is why run count, not prompt count, is the number to argue about.
One honest caveat: runs are not independent. The same prompt asked twice, or asked across two engines that draw on overlapping sources, produces correlated answers. DerivateX inflates the variance by a design effect of 2, which widens the interval by a factor of the square root of 2, giving roughly plus or minus 3.4 points at 1,200 runs. We chose that multiplier as a conservative convention and we say so in the footnote, because a stated assumption someone can challenge is worth more than a clean number nobody can audit.
Five, and DerivateX will not sign off on an AI search measurement model that is missing any of them. Prompt logs give you exposure. Analytics gives you arrival. The CRM gives you pipeline. Self-reported attribution catches the discovery that leaves no referrer. Third-party citation data tells you which surfaces to work on next.
| Layer | Metric | Formula | Source | Cadence |
|---|---|---|---|---|
| Exposure | Recommendation share | Vendor-slot runs / total runs | Prompt log, fixed frame | Weekly, reported monthly |
| Exposure | Owned citation share | Own-domain citations / total citations | Prompt log citation capture | Weekly |
| Exposure | Influenced citation share | Third-party citations naming brand / total citations | Prompt log plus manual page check | Monthly |
| Exposure | Prompt coverage | Prompts with any brand presence / total prompts | Prompt log | Monthly |
| Arrival | AI referral sessions | Sessions with AI source or UTM | GA4, server logs | Weekly |
| Arrival | Branded search lift | Branded impressions vs 90-day baseline | Search Console | Monthly |
| Revenue | AI-sourced pipeline | Opportunity value where first touch is an AI surface | CRM | Monthly |
| Revenue | AI-assisted pipeline | Opportunity value with an AI touch anywhere in path | CRM | Monthly |
| Revenue | AI-attested pipeline | Opportunities where the buyer names an AI tool, no click evidence | Form field, coded | Monthly |
| DerivateX locks all five layers into the measurement doc at kickoff, so every reported figure can be recomputed from the raw logs by someone else. | ||||
On the arrival layer, OpenAI documents ChatGPT search referrals arriving with utm_source=chatgpt.com in its publishers and developers FAQ, which makes a clean GA4 segment possible without guesswork. Build that segment on GA4 traffic-source dimensions rather than on a channel grouping you have not inspected, and treat the result as a floor on ChatGPT-driven arrival, never as the total, because in-answer recommendations that produce no click will never appear there. Google separately documents how AI features in Search surface site content in its Search Central guidance, which is worth reading before anyone argues that Google AI Overviews can be blocked or forced.
On the revenue layer, you need one required field on the demo form asking how the buyer first heard about you, with free text, coded weekly. Most SaaS teams do not have this instrumented at all. It lives in your CRM of record, whether that is HubSpot, Salesforce or a lighter system such as Pipeline CRM, which lists 25 custom fields on its Start plan and unlimited fields on Grow at $49 per user per month billed annually, per its published pricing. The field limit is a real constraint on smaller stacks, so check it before designing the schema.
Here is the worked example DerivateX uses to teach the model. A B2B SaaS company at roughly $18M ARR, one category, four engines, 150 prompts, two repeats, so 1,200 runs per cycle and 5,400 citations captured.
| Metric | Baseline | Day 90 | Change | Outside noise? |
|---|---|---|---|---|
| Recommendation share (unweighted) | 288 / 1,200 = 24.0% | 396 / 1,200 = 33.0% | +9.0 pts | Yes, interval is plus or minus 3.4 pts |
| Recommendation share (weighted) | 19.1% | 27.4% | +8.3 pts | Yes |
| Owned citation share (link-level) | 189 / 5,400 = 3.5% | 243 / 5,400 = 4.5% | +1.0 pt | No, inside the interval |
| Influenced citation share | 810 / 5,400 = 15.0% | 1,404 / 5,400 = 26.0% | +11.0 pts | Yes |
| Prompt coverage | 61 / 150 = 40.7% | 82 / 150 = 54.7% | +14.0 pts | Yes |
Read the table the way an operator would. Owned citation share barely moved. Influenced citation share moved 11 points, and recommendation share moved 9, so the recommendation gain came from third-party corroboration rather than from the company’s own pages. That is the normal pattern, and it is consistent with 46.7% of Perplexity top sources coming from Reddit and 80% of URLs cited by ChatGPT and Perplexity sitting outside Google’s top 100. A team reporting only owned citation share would have concluded the quarter failed.
The revenue columns sit in a separate view, deliberately. Same cycle for the same illustrative company: 41 AI-sourced opportunities, 96 AI-assisted, 63 AI-attested with no click evidence. Do not add attested to sourced. Report them adjacent, note the overlap risk, and let leadership see the range instead of a single confident figure.
To be explicit, every figure above is a teaching example and none of it comes from a DerivateX account. On real client work the same discipline is what makes a published number stand up: Gumlet attributes more than 20% of monthly inbound revenue to AI discovery, and that statement holds because the attribution logic was defined before the number was quoted.
You separate the claim into three tiers and refuse to merge them. DerivateX writes the tier definitions into the reporting spec so that nobody downstream can quietly promote an attested deal into a sourced one at quarter end.
The test DerivateX applies before any causal sentence goes into a deck is a lag test. Did recommendation share rise in a specific prompt cluster, and did attested or sourced opportunities mentioning that use case rise 30 to 60 days later? If both are true and no other channel changed materially in the same window, you have a directional argument. If only the second is true, you have a coincidence and should label it as one.
| These metrics can prove | These metrics cannot prove |
|---|---|
| Whether your brand is present in a fixed, repeatable prompt frame | What any individual buyer actually saw in their own session |
| Directional movement over time, when the frame is locked and runs exceed the noise floor | The absolute size of your AI-driven demand |
| Which sources engines rely on in your category | That a citation caused a recommendation |
| Relative standing against a named competitor set on identical prompts | Market share, or share of all prompts ever asked in your category |
| Which prompt clusters are uncovered, and therefore where to work next | That a recommendation share gain caused a specific closed deal |
DerivateX puts a version of this table in client reporting because 64% of marketing leaders are unsure how to measure AI search, and the fastest way to lose credibility with a board is to overclaim once. The same logic applies to claims about our own work: Verito moved from an average position of 40 on Google to first page, and is cited and recommended on ChatGPT and Google AI Overviews for 40 of their commercial hosting queries. That is a countable statement about a defined query set, not a claim about the whole category.
One slide, three rows, and a footnote. DerivateX builds the leadership view as exposure, arrival, revenue, in that order, with the interval printed next to every share figure and the prompt frame version number in the footer.
Two internal frameworks carry this once the reader has the mechanics. The Citation Surface Map is the DerivateX inventory of every domain that engines cite in a client’s category, ranked by citation frequency and by whether the client can influence it, which is what turns influenced citation share from an observation into a work queue. The AI Visibility Score, or AVS, is the composite DerivateX index that combines weighted recommendation share, prompt coverage and citation share into a single trackable figure for executives who want one line rather than nine. The composite is for the board slide, the component metrics are what the team works from, and the composite never replaces them in the working report.
The underlying method is Citation Engineering, which is the DerivateX methodology for building and placing the corroborating evidence that makes language models recommend a brand deliberately rather than incidentally. The DerivateX B2B SaaS AI citation study sets out which source types recur across categories, and our breakdown of ChatGPT software recommendation sources covers where the vendor slots are actually drawn from.
When you have an analyst who will run the frame and someone who will act on the output. DerivateX is the wrong purchase for a team that mainly needs visibility into the numbers, and saying otherwise would waste both sides three months.
Similarweb’s AI Search Intelligence is a genuinely capable option here, and on breadth of surface coverage it does something a founder-led agency does not try to replicate. It reports AI Brand Visibility as the percentage of AI responses that mention your brand out of all analyzed answers, offers prompt tracking, citation analysis of the domains most influencing answers in your category, sentiment breakdowns and AI traffic across ChatGPT, Perplexity, Gemini and AI Mode. The pricing FAQ on that same page lists a standalone AI Search Intelligence tier at $99 and a second no-touch tier at $333, with no publication date shown, so treat both figures as current at the time of writing and check the page before you budget. It also states that Similarweb is trusted by 200-plus of the Fortune 500. Note that a mention-based visibility percentage is not the same numerator as a vendor-slot recommendation share, so if you adopt a tool, read its definition and match your reporting language to it. Our comparison of SEO and GEO tools for B2B SaaS goes deeper on how the vendors differ in what they count.
DerivateX publishes its pricing in full, quoted all in as retainer plus off-site budget rather than as a retainer alone.
| Your situation | Best purchase | Typical cost |
|---|---|---|
| You need the numbers and have an analyst to run them | A tracking tool, run in house | Tens to low hundreds per month |
| You need to know whether there is a problem before you commit budget | DerivateX Diagnostic, two weeks, credits in full against month one if converted within 30 days | $3,500 one time |
| You need the frame built, run and acted on across one category | DerivateX Rank and Get Found, 90-day pilot, no lock-in after | $6,000 to $6,200 all in per month |
| You need multi-category coverage and board-grade attribution | DerivateX Own Your Category, 90-day pilot | $9,500 to $10,000 all in per month |
DerivateX commits to the cadence, the frame integrity and the measurement, with targets rather than guarantees on the share figures themselves. What a defined frame makes possible is a claim you can check: REsimpli became the most cited and recommended real estate CRM for investors in ChatGPT within 90 days, and that statement is only meaningful because the prompt set behind it was fixed in advance.
Fix a prompt frame of 120 to 200 buyer prompts, run each twice per engine, then divide vendor-slot appearances by total runs for recommendation share and own-domain citations by total citations for citation share. DerivateX reports both with a 95% interval attached so leadership can tell signal from sampling noise.
There is no published benchmark DerivateX would stand behind. With 3 to 4 brands recommended per AI category query, the arithmetic ceiling for any single vendor is high but slot competition is fierce. Set targets against your own locked baseline and a named competitor set on identical prompts instead.
Partially, yes. OpenAI documents that ChatGPT search referrals can arrive carrying utm_source=chatgpt.com, which makes a clean GA4 segment possible. DerivateX treats that segment as a floor, not a total, because recommendations that produce no click, or that produce a branded search days later, will never appear in it.
Yes, as the exposure row above arrival and revenue, with the confidence interval and prompt frame version printed alongside. DerivateX pairs it with sourced, assisted and attested pipeline in separate columns, because merging attested into sourced is the single fastest way to lose credibility at the next review.
Citation share usually moves first, within weeks of new corroborating evidence being indexed. Recommendation share typically follows, and DerivateX plans engagements on a 90-day measurement cycle for that reason. Pipeline effects appear later still, which is why the lag between each layer is measured rather than assumed.
Citation share is your supply of evidence. Recommendation share is your position in the answer. Pipeline is what happens when the position holds long enough for buyers to act on it. Work on the first, report the second, and prove the third separately with CRM logic you would be comfortable defending to a finance team. In DerivateX’s experience, software firms that collapse all three into one number tend to end up with a metric that rises while pipeline does not, and no way to explain why. Traffic and rankings were always inputs, and share metrics are inputs too. The output is a shortlist you appear on when a buyer asks a machine what to buy.
Request the free AI visibility audit from DerivateX at derivatex.agency/free-ai-visibility-audit and you get your recommendation share and citation share baseline across ChatGPT, Perplexity, Gemini and Google AI Overviews back within 48 hours.
Apoorv Sharma is the co-founder of DerivateX, a B2B SaaS SEO and Generative Engine Optimization agency that engineers AI citations in ChatGPT, Perplexity, Claude, and Gemini and connects them to demo bookings and revenue pipeline. He is the author of the 2026 AI Visibility Benchmark Report and the Citation Engineering methodology. He's also the brain behind "Found On AI" and has sold 2 of his companies previously
Shivanshi Bhatia is the co-founder of DerivateX, a B2B SaaS SEO and Generative Engine Optimization agency that engineers AI citations in ChatGPT, Perplexity, Claude, and Gemini and connects them to demo bookings and revenue pipeline. She runs operations and delivery, which means every audit, content brief, and published page ships through a system she built. She owns the client relationship from kickoff through reporting, so clients spend their time on decisions instead of chasing updates. She has worked in SaaS since 2019 and reviews client work before it goes live.
