To check which brands ChatGPT recommends in your category, one screenshot of one answer won’t get you far. You need a fixed set of questions, run the same way each time and coded against rules you wrote down before you looked. This is the method behind the High Salience teardowns, cut down to steps, with the CRM teardown as the running example.
Step 1: Fix your brand list before you collect
Fix your brand list before you collect a single answer. List every brand a buyer could plausibly be sent to, with its aliases and its own domain, so “monday”, “monday.com” and “monday CRM” count as one brand and you know which sites are first-party. The CRM run used a 52-brand dictionary, frozen before collection. Build the list after reading answers and you end up counting whatever caught your eye.
Brands outside the list still matter. Note them by hand, and flag any answer that recommends only outside brands as having left your competitive set. In CRM, the real estate and law firm questions landed there.
Step 2: Write 100 queries from four templates
Write your 100 queries from four templates so the mix of intents stays the same every time you run the test. High Salience splits them 30 category (“best CRM for small teams”), 30 comparison (“salesforce vs hubspot”), 20 alternative (“cheaper alternative to hubspot”) and 20 recommendation (“which CRM should I use for a 5 person sales team”).
Work a few vertical versions into the category and recommendation sets. In CRM, a single word like “real estate” was enough to swap out the whole field. Freeze the list and publish it with your results, the way the CRM query list is published, so anyone can rerun it.
Step 3: Run each query once in logged-out ChatGPT with web search on
Run each query once in logged-out ChatGPT with web search on, from one country and in one language, and log the collection date. High Salience collects through the DataForSEO scraper with web search forced on, in US English. If you run the queries by hand, a logged-out session keeps your own chat history out of the answers.
One run per query keeps the job small. It also leaves variance unmeasured, so say that wherever you publish. That gap is why High Salience’s benchmark repeats every query three times on each of three surfaces.
Step 4: Save the full answer and every citation
Save the full answer and every citation, not your summary of them. Keep the text with its tables and every cited URL. When you code, ignore link addresses and read only the words a buyer sees. The published codebook shows why: under an older rule, a citation whose address contained a brand name, like “alternatives-to-salesforce”, counted as a mention of that brand.
For each mention, the CRM dataset also records its position in the answer, whether it was bolded or sat in a table, and whether the brand was named in the query. Later, collapse citations to the domain, one per domain per answer, and tag each one first-party or third-party.
Step 5: Code each brand mention with one fixed label
Code each brand mention with one fixed label per answer, and let the strongest label win when a brand comes up more than once. These are the labels in High Salience’s reader protocol:
- Picked: the answer’s own verdict, overall or for a case, such as “my pick,” “choose X if,” a shortlist it tells the reader to try, or the top of its own ranking.
- Recommended: tied to a stated fit without being the verdict, like “X: best for agencies” in a list of eight options.
- Listed: offered as an option but not matched to any buyer or case.
- Passing: named but not offered, such as an integration, a comparison anchor or a source label.
- Anchor: the brand an alternatives query is trying to replace. Keep it out of every total.
When picked and recommended blur, ask whether the answer is telling the reader what to choose or describing who each option suits. Throw out false matches too. “Close probability” is not the CRM called Close.
Step 6: Use two coders and a third to settle disagreements
Use two coders who work separately, plus a third to settle the codes they disagree on. In the CRM teardown the coders were two AI readers (Claude agents) working from a written protocol, and they agreed on 98.6 percent of brand codes, 99.3 percent at the recommended level. Publish your agreement rate, whatever it turns out to be.
Don’t trust a keyword script on its own. The keyword-rule coder that produced the first CRM figures agreed with the readers on only 70.5 percent of brand codes. And if both coders are AI models, a blind spot they share won’t show up as a disagreement.
Step 7: Pull Google’s top ten as a control
Pull Google’s top ten for the same queries, collected in the same window, as your control. Request 20 results and keep the first ten organic ones; the codebook notes that a request for ten came back with only six to nine organic rows once ads, videos and People Also Ask took their slots.
For each recommended brand, record whether its own domain is in that top ten. A listicle that mentions the brand is not the brand ranking. Do the same for every cited domain. In CRM, 269 of the 387 recommended brand mentions were for brands whose own site missed the top ten.
Check the control itself before you trust it. Since late September 2026, Google has answered some automated searches with a swapped page full of the wrong sites, and 481 of the 1,400 Google pages I pulled for High Salience on September 26 came back that way. Open a handful of your searches in a browser and compare the sites, the same check I ran on two data sources in my Cloro vs DataForSEO test.
Step 8: Put the denominator in every sentence
Put the denominator in every sentence you write about the results. “HubSpot was picked in 69 of 100 answers” can be checked, while “HubSpot dominates” can’t. Report three numbers per brand (appears in, recommended in, picked in), each out of the same 100 answers.
Give each metric one definition and keep it. A pick rate is picks divided by appearances (Pipedrive’s was 53 of 55), and recommended brand mentions are mentions, not brands. Count citations once per domain per answer, so a site linked three times in one answer doesn’t triple its weight. Then check that your counts add up before anything goes out.
Step 9: Write the limits before the findings
Write the limits before the findings, so they shape how you word everything else. One run per query leaves variance unmeasured. Two categories collected on different days can’t be compared without the date muddying it. A brand you didn’t see was not observed in your sample, which is a long way from zero visibility. If your dictionary leaves out vertical tools, say so, since the vertical questions are where those tools turn up.
What the CRM run turned up
What the CRM run turned up shows the kind of result this method produces. HubSpot was picked in 69 of the 100 answers, every Salesforce recommendation came with a big-company condition, and two real estate questions pushed all 52 dictionary brands out. The full numbers are in my breakdown of the CRM teardown.
Sources
- 100 CRM Questions: What ChatGPT Recommends (High Salience)
- CRM teardown codebook, v2.0 (High Salience)
- Reader protocol for brand coding (High Salience)
- CRM teardown query list (High Salience)