Your pricing page can help an AI recommend your competitor. If you market a business whose buyers ask ChatGPT or Claude what to buy, here are the two contenders I would keep separate: brand-recommendation share of voice and owned-source citation coverage. One measures whether you make the shortlist. The other measures whether your website supplies the evidence. Blend them and you hide the decision that matters: fix the source material, or fix your brand’s place in the answer.
Say you spend $20k/month and someone asks ChatGPT for the best option in your category. The answer lists three competitors. Underneath sits a link to your pricing page. Your tracker counts a win because you got cited. The buyer never even considers you. Expensive footnote.
The reverse happens too: the answer recommends your business, but every supporting link points to Reddit, a review site, or a publisher. You win the recommendation without owning the cited evidence. In one analysis of 3,981 domain appearances, 61.7% were cited but never named, 25.1% were mentioned without a citation, and only 13.2% were both.
These scores are not interchangeable. I would compare them on four things: what they count, what they miss, how reliably you can measure them, and which work they tell you to do.
1. What they count: a place on the shortlist vs a place in the sources
Brand-recommendation share of voice counts buying consideration
Take 20 to 30 prompts your best customers would type. Run them in the assistants those buyers use. Count the answers where your brand is named as a pick, then divide by all answers checked.
Not mentioned in passing. Recommended.
Before calculating anything, record the scope:
- The exact prompt list.
- The competitors you are tracking alongside your brand.
- The assistants and versions, where available.
- The capture dates.
I used to tell clients to track five prompts in ChatGPT and call it share of voice. I was wrong. Commercial trigger words like deals, where, and cheap drive 4 to 8x higher mention rates. A prompt set full of best and cheapest can tell a very different story from an informational set. Know which contest you are measuring before declaring a winner.
Owned-source citation coverage counts your website’s presence
Use the same answers, but change the numerator: count answers with at least one cited URL you control, then divide by all answers checked.
Pull every cited link and match its domain against your own. Include subdomains and help docs. Exclude the review site that happens to host your logo. That is someone else’s source, however flattering the page looks.
The distinction becomes obvious with a buying prompt such as best project tool for a 20-person agency that bills hourly. The answer can recommend Asana, Monday, and ClickUp while citing your pricing page for one fact. Your domain appears. Your business does not make the shortlist.
My rule: count recommendations and owned citations separately, on the same answers. Neither score gets to stand in for the other.
2. What they miss: borrowed authority vs footnote status
Recommendation share is closer to the buying decision, but it cannot tell you whether you own any of the proof.
You can win 40% recommendation share while every citation underneath points to G2, Reddit, or a publisher. I call that borrowed authority. Those pages support your presence, but you do not control them. If a third party rewrites its comparison, your recommendation can disappear with it.
Citation coverage has the opposite blind spot. It tells you that your source appears, not that the answer endorses your business. A pricing page can supply the cleanest statement of your price while a competitor takes the recommendation sentence.
BrightEdge found ChatGPT mentions brands 3.2x more often than it cites them, and fewer than half of the top-100 mentioned brands also rank in the top-100 cited domains. Those are mention measures, not recommendation measures. That is another reason to inspect the answer rather than trust the label on a chart.
Even a citation-only improvement can look more commercially useful than it is. In Pew browsing data covering 68,879 Google searches, users clicked a traditional result on 8% of visits with an AI summary versus 15% without, and clicked a link inside the summary on only 1% of visits. A source link is not automatically a visit, much less a reason to buy.
To inspect the split, paste five to ten saved answers into Claude and ask it to separate:
- Every named brand.
- Every recommended brand.
- Every cited URL.
- Answers where your domain appears but your brand is not recommended.
For a closer walkthrough, see why ChatGPT can cite your website while recommending a competitor.
Ask what sentence your page supported. If it supplied a price or definition, count the citation. Do not promote it to an endorsement.
3. How to measure them: identical answers, visible evidence
Run both counters on the same saved answers or the comparison is theater. Same prompt list. Same capture window. Same assistants, with versions recorded where available. ChatGPT today is not ChatGPT next month.
The full answer matters because a yes/no label hides the distinction you are trying to measure. With groas collecting the answers, I would give Claude two separate instructions:
Show which tracked answers recommend us and which recommend competitors. Preserve the full answer and capture date.
Then:
Show answers citing our website, including answers that never recommend our brand.
That second clause does real work. Without it, a report can quietly turn source coverage into commercial success.
Keep missing data separate, too. If an assistant is down, a prompt times out, or the sources were not saved, mark the record missing. Missing is not zero. Zero means the answer ran and you were absent. Treat a hole in the log as zero and you manufacture a crisis out of a broken scrape.
Any dashboard worth paying for should preserve:
- The prompt and capture date.
- The full answer text.
- The recommended brands.
- The cited URLs.
- Missing or incomplete captures.
A tracker that shows 72 and hides the proof gives you nothing to fix. groas collects answers and sources around the clock through its AI search data collection, so Claude can re-read the evidence after a page change instead of asking you to trust a smiley face.
For the full counting routine, use this walkthrough to track AI brand mentions across ChatGPT, Claude, and Perplexity.
Save the evidence first. The score comes second.
4. Which work they trigger: brand positioning vs source material
When recommendation share is weak, I start with the pages that describe the brand’s fit for the buying job: reviews, comparisons, Reddit threads, and publisher coverage. What category words do they use? Which competitors do they describe as suitable? Where does your brand disappear from that conversation?
There is a reason to examine those mentions. Ahrefs found branded web mentions correlate 0.664 with AI Overview visibility across 75,000 brands, compared with 0.218 for backlinks. That is not a recommendation-share score, but it points toward evidence worth inspecting rather than another round of polishing your own homepage.
When citation coverage is weak, I start with the owned page. Does it give the answer a clear fact to use? Is that fact available in crawlable HTML? Does the page state the price, feature, or definition plainly, or bury it under three paragraphs of positioning language?
In one distributed-content test, brand pages were cited in 13.6% of responses but named in only 3.8%. That gap is exactly why I would not use a citation increase as proof that brand positioning is fixed.
The work overlaps, but the diagnosis should not. A clearer pricing fact addresses a different gap from a comparison that never considers your business.
Recommendation weakness sends me to brand fit and supporting mentions. Citation weakness sends me to the source page. Here is the head-to-head in one view:
| Decision axis | Brand-recommendation share of voice | Owned-source citation coverage |
|---|---|---|
| What it counts | Answers where your brand is named as a pick | Answers with at least one cited URL you control |
| What it misses | Whether you own the cited evidence | Whether your business makes the shortlist |
| Measurement requirement | Full answers, fixed prompts, assistants, and capture window | The same answers and window, with cited URLs preserved |
| Work it triggers | Inspect reviews, comparisons, Reddit, and publisher mentions | Inspect page clarity, pricing, and quotable facts |
| Risk when used alone | Borrowed authority you do not control | Footnote status without buying consideration |
The verdict: recommendation share wins for businesses selling on recommendation
For a business selling on recommendation, brand-recommendation share of voice wins. It measures the contest that puts you on the shortlist. Citation coverage is the control panel, not the prize.
Who should lead with the other score? A docs-led developer tool, a pricing-transparency brand, or a site that gets paid when a fact gets pulled into an answer. For those readers, owned-source citation coverage deserves the lead position. For everyone else selling on recommendation, fix recommendation first.
That does not mean ignoring the second score. It means using both to choose the work, rather than averaging away the problem. Anyone selling you a single blended visibility number is selling you a shortcut around this decision.
Recommended and cited: protect the evidence
Do not touch what is working until you know why. Save the answers. Note which third-party pages support the recommendation and which owned page earns the citation. Protect both.
Re-check the same prompt list in two to four weeks. This is the quadrant where leaving it alone can be the correct action. Log it before someone commissions a rewrite to look busy.
Recommended but not cited: inspect the borrowed authority
Buyers meet your business through G2, a Reddit thread, or a publisher roundup, but your own site never appears underneath. Find the exact pages doing the work, then build the owned page that deserves to support the same claim: the same buying language, backed by your evidence.
You are not chasing a citation for the chart. You are reducing reliance on pages someone else can edit. Keep the recommendation while strengthening the source material you control.
Cited but not recommended: inspect the buying answer
Your URL sits in the footnotes while your brand misses the shortlist. Open the cited page and read the first screen like a buyer. Does it answer the job in the prompt, with a price, a fit statement, and proof? Or does it define a term and stop?
If the page only supplies a fact, rewrite the top 150 words to state who it is for and why it wins. Keep the fact that earned the citation. Then inspect the third-party pages supporting the competitors’ recommendations.
Fix the missing buying argument without destroying the useful source. A footnote and a shortlist place are different jobs.
Neither: find the gap before ordering content
Do not order five blog posts. Pull the three competitor answers that keep winning your prompts and list the domains cited underneath. Identify one gap on one page that serves that prompt.
Start with the evidence already winning the answer. A new article no retrieval system asked for changes nothing.
Turn the losing quadrant into one approved task in Claude
Pick the quadrant costing you the most pipeline and turn it into a single change. Ask Claude, with groas connected, for the supporting pages behind your losing answers and the SEO data on those pages. Approve one edit: a clearer fit statement, a pricing fact, or a missing comparison row.
Log the reason, publish, and re-check the same prompts in two to four weeks.
Trackers can display mentions, citations, and share of voice together across ChatGPT, Perplexity, Gemini, Claude, and AI Overviews. That display is not a diagnosis. The diagnosis is this sentence: We were cited without being recommended, so we changed this page for this prompt, on this date.
Run the two scores again on the same measurement setup. If recommendation share moves, you changed your presence in the answer. If only citations move, you changed the footnotes. Do the first one on purpose.

