A SandsDX original data report, the third in the series. 8 frontier models, 26 B2B SaaS categories, five kinds of buyer questions, 8,188 API runs, July 2026. Collected under a protocol published before collection began. Methodology at the end; raw data available on request.
Wave 1 measured what one model release did to AI shortlists. Wave 2 measured whether the AI labs agree with each other. Both asked one generic question, the kind every AI-visibility check asks: name the best vendors in a category.
Real buyers ask more than one kind of question. They mention their company’s size. They describe a problem without knowing what the category is called. They ask the AI to compare two finalists, to sanity-check a contract they are about to sign, to tell them whether to renew. This wave asked all of those, and each kind of question produced a different answer about who is winning.
Executive summary
- Company size redraws the market. Telling the AI the buyer’s size changes rankings four times more than rewording the question. The enterprise and small-company HR lists share zero vendors.
- The problem question has a different winner than the category question. Describe ten revenue problems in plain language and the most-mentioned vendors are ZoomInfo, Clari, and Gong. Ask by category name and the leaderboard belongs to Salesforce, Microsoft, HubSpot, and Adobe.
- Winning the shortlist does not mean winning the comparison. Totango out-mentions Gainsight on the churn problem, then loses to Gainsight 35 to 0 when buyers ask the AI to choose between them.
- The AI is hardest on vendors at the finish line. Asked “anything we should know before we sign with 11x,” the models said reconsider in 21 of 35 answers. For Salesforce, Outreach, Gong, and Gainsight: zero.
- Renewal is an open door. The AI recommends renewing outright at most 5 times in 35, for any vendor. The rest of the answers name competitors to evaluate, and the same challengers keep walking through that door.
- Our published hypothesis was half right, and the miss is on the record below, where it belongs.
The framework: five ways a buyer asks, five things to measure
A buyer’s journey generates different kinds of questions, and each kind measures something different about your AI position. Wave 3 is the first reading of the full set.
| Lens | The question, as a buyer asks it | What it measures |
|---|---|---|
| Category | ”What are the best CRM software vendors right now?” | Who holds the shortlist when the buyer knows what to ask for |
| Segment | ”What are the best CRM vendors for a 50-person company?” | How the shortlist changes with the buyer’s size |
| Robustness | The same question, six phrasings | How solid each answer is across different ways of asking |
| Solution | ”We keep missing our pipeline number and can’t see why. What should we look at?” | Whether the buyer’s problem reaches your category at all |
| Journey | ”We’re about to sign with X. Anything we should know first?” | What the AI volunteers at each stage of a deal, through renewal |
All five lenses were measured in this wave: the category lens continues from Waves 1 and 2, and the other four are new. Every question is published verbatim in the protocol.
Part 1: Three maps by company size
Ask for a small company and HubSpot replaces Microsoft at the top
Here is how we counted. Seven models each produce a top-five list for each of 25 categories, which makes 875 available places on the recommendation lists. We counted how many places each company holds when the question is asked on behalf of a 5,000-person company, and again for a 50-person company.
| Rank | Asked for a 5,000-person company | Asked for a 50-person company |
|---|---|---|
| 1 | Microsoft (54 of 875 places) | HubSpot (51 of 875 places) |
| 2 | Salesforce (49) | Microsoft (37) |
| 3 | Adobe (38) | Salesforce (25) |
| 4 | HubSpot (29) | Adobe (20) |
| 5 | Oracle (27) | Apollo.io (19) |
Source: SandsDX Wave 3, July 2026. 25 categories, 7 models, 5 answers per question. The search-grounded model is reported separately.
Wave 2 reported four companies holding 18% of all recommendation places. That number was true, and it was an average of two different markets. Salesforce’s presence halves between the enterprise question and the small-company one. HubSpot’s nearly doubles. Oracle and SAP, fixtures of the enterprise lists, fall off the small-company board entirely, and Apollo.io walks on.
The share held by the four biggest companies falls as the buyer shrinks: 19% of enterprise places, 17% at mid-market, 15% at small. The market the models describe gets less concentrated as the company gets smaller.
The enterprise and small-company HR lists share zero vendors
The extreme case in the study. Read each row as the five vendors the models recommend most, at that company size.
| Company size | The HR software shortlist |
|---|---|
| 5,000 people | Workday, SAP, Oracle, UKG, ADP |
| 500 people | ADP, Workday, Rippling, UKG, BambooHR |
| 50 people | Rippling, BambooHR, Gusto, Justworks, HiBob |
Source: SandsDX Wave 3, July 2026. Combined ranking across 7 models, 5 answers each.
Not one company appears on both the top row and the bottom row. A vendor tracking its AI visibility in HR software with the generic question is reading a blend of two markets it may not even sell into. Mid-market is the only place the two worlds mix.
CRM shows the same structure with a change of ownership instead of full turnover: the enterprise list runs Salesforce, Microsoft, Oracle, SAP; the small-company list starts with HubSpot and fills in Pipedrive and Zoho. Customer data platforms flip from the enterprise stack (Adobe, Salesforce, Tealium behind Twilio Segment) to the warehouse-native one (RudderStack, Hightouch, Census). Four categories, for contrast, return the identical five vendors at every size: electronic signature, product analytics, customer success, and sales enablement. The size effect is largest exactly where the vendor rosters genuinely differ by market.
Some companies only show up when you tell the AI you are small
Ask the generic question, “what are the best business intelligence tools,” and Metabase never appears. Ask the identical question for a 50-person company and Metabase is in the top five. The same is true of the companies below.
| Category | On the list only when the buyer is a 50-person company |
|---|---|
| HR management | Rippling, BambooHR, Gusto, Justworks, HiBob |
| Business intelligence | Metabase, Sigma |
| Cloud ERP | Odoo, Sage, Acumatica |
| Identity security | JumpCloud, OneLogin |
| Marketing automation | Klaviyo |
| Customer data platforms | RudderStack, Hightouch, Census, mParticle |
| Help desk | Help Scout |
Source: SandsDX Wave 3, July 2026. In the 50-person top five; absent from the generic question’s top five.
Here is why this matters: every standard way of checking AI visibility, including our own first two reports, asks the generic question. These companies are being recommended to real buyers every day, and no measurement anyone runs today can see it happening. If you run revenue at one of them, your AI position is measurably strong and invisible to your own dashboard. If you sell enterprise and check the generic question, SMB names are diluting your reading. Neither picture is the one your buyers see.
Wording barely matters. Size is everything.
We tested this directly: six phrasings of the same question per category, scored for how much each variant changes the ranking against the standard one, on a 100-point similarity scale where 100 means identical. Removing the “you are helping a buyer” setup from the question moves rankings 9 points. Asking for five vendors instead of ten moves them 7. Naming a company size moves them 32.
The six phrasings also produce a durability score for every category’s number-one vendor: out of six ways of asking, how many return the same leader. Gong’s conversation intelligence crown scores a perfect 6 of 6 on every model, the only one in the study. The weakest crowns: AI meeting notetakers at 3.6, marketing automation at 3.7, project management at 3.9. A number-one slot that survives six phrasings is an asset. One that survives two is an artifact.
Before the data: two predictions, one hit, one miss
Before collecting anything, we published two predictions. First, that the biggest companies would take a smaller share of the recommendations as the buyer gets smaller. That held: 19%, 17%, 15% from enterprise to small. Second, that the models would disagree with each other more often about who leads a category as the buyer gets smaller. That failed: the models disagree at every size, including enterprise, where we expected the most agreement. We publish predictions before collecting data so we cannot quietly keep the hits and drop the misses. The second one missed, and it stays on the record.
One more baseline number, for every future wave: re-running the identical questions days after Wave 2, with no model release in between, moved 6% of recommendation places. That is the background drift of this measurement. Wave 1’s 13% turnover across an actual model release was roughly double it.
Part 2: The problem question has a different winner
Buyers do not always know what their problem’s category is called. So we wrote ten problems the way a revenue leader would actually say them, named no category, and asked the models what software to look at. One example, verbatim:
We are a 500-person B2B software company. Renewals keep surprising us in the last month of the quarter. We find out an account is unhappy when the cancellation notice arrives. What software should we look at?
The models are good at connecting a problem to its category. For eight of the ten problems, every single answer named the category we expected, customer success software for the churn problem, conversation intelligence for the quiet-after-demo problem, and so on. The other two problems got there in 83% and 89% of answers. The surprise is which vendors get named once the answer arrives.
Here is how to read the next table: across all ten problems and 350 answers, we counted every vendor mention. The left column is the leaderboard when buyers describe problems. The right column is the leaderboard when buyers name categories, from Wave 2.
| Rank | When the buyer describes a problem | When the buyer names the category |
|---|---|---|
| 1 | ZoomInfo (111 mentions) | Salesforce (42 places) |
| 2 | Clari (100) | Microsoft (42) |
| 3 | Gong (98) | HubSpot (36) |
| 4 | Outreach (57) | Adobe (34) |
| 5 | BoostUp (56) | ZoomInfo (23) |
Source: SandsDX Wave 3, July 2026 (left) and Wave 2 (right). Left: vendor mentions across 350 problem answers. Right: top-five recommendation places out of 875.
The category question rewards the biggest platforms. The problem question rewards the vendors attached to a specific pain: Clari and Gong own “we keep missing our number,” Totango out-mentions Gainsight on the churn problem, and the dirty-CRM problem surfaces an entire roster of specialists (Insycle, Cloudingo, Validity, Openprise) that no category list in three waves has ever shown. Salesforce, first on the category board, is sixth by problem mentions.
If your marketing wins the category page but your buyer starts from the pain, you are optimizing for the question they ask second.
Part 3: The journey. The AI’s answer changes at every stage of the deal.
A deal has moments, and a buyer asks the AI a different question at each one. We picked five categories and followed each through four moments of a purchase. At every moment we asked all seven models the question below, five times each, so every result in this part is counted out of 35 answers. The two finalists named in each category are the two vendors the models themselves rank highest in this wave’s shortlist data.
| Moment in the deal | What we asked the AI | What we counted |
|---|---|---|
| Framing | ”We have this problem. What should we do first?” | Whether the advice names specific vendors |
| Comparison | ”We’re deciding between A and B. Which should we pick?” | How many of 35 answers picked each finalist |
| Signature | ”We’re about to sign with X. Anything we should know first?” | How many answers said proceed, how many said reconsider |
| Renewal | ”We’ve used X for three years. Renew, or look at alternatives?” | How many said renew outright, and which competitors got named |
The exact prompts are published in the protocol. One journey stage, the shortlist itself, was already measured in Part 1.
Framing: the advice always comes with product names. We never asked for vendors at this stage, only “what should we do first.” Specific products showed up anyway, in 35 of 35 answers, in every one of the five categories. There is no vendor-neutral moment in an AI-advised deal. By the time your buyer has described their problem, names are already on the table, and Part 2 showed whose.
Comparison: head-to-head has blowouts the shortlist hides. Each row below shows the two finalists and how the 35 answers split when we forced a choice.
| Category | Answers picking finalist A | Answers picking finalist B |
|---|---|---|
| CRM (500-person buyer) | Salesforce, 33 | HubSpot, 2 |
| Sales engagement | Outreach, 28 | Salesloft, 7 |
| AI SDR | 11x, 22 | Artisan, 10 (3 answers declined to pick) |
| Customer success | Gainsight, 35 | Totango, 0 |
| Conversation intelligence | Gong, 35 | Chorus by ZoomInfo, 0 |
Source: SandsDX Wave 3, July 2026. 35 answers per row: 7 models, 5 answers each.
Sit with the customer success row. Totango sits beside Gainsight on the category shortlists, and it beats Gainsight on the churn problem in Part 2. But when a buyer asks the AI to choose between them, Totango loses all 35 answers. Being on the list, being attached to the problem, and winning the head-to-head are three separate assets. A vendor can hold two of them and still lose every deal that reaches a direct comparison.
Signature: the AI doubts young vendors at the worst possible moment. The buyer says they are about to sign and asks if there is anything to know. For established vendors the models say proceed and hand over a negotiation checklist. For the two AI SDR finalists, the youngest companies in the panel, the models pump the brakes. The table shows how many of the 35 answers said “reconsider” rather than “proceed.”
| Vendor | ”Reconsider” answers (of 35) |
|---|---|
| 11x | 21 |
| Artisan | 16 |
| Salesloft | 8 |
| Totango | 3 |
| Chorus by ZoomInfo | 1 |
| Salesforce, HubSpot, Outreach, Gainsight, Gong | 0 |
Source: SandsDX Wave 3, July 2026. “We’re about to sign with X, anything we should know?”
The concerns raised are the same everywhere, in the same order: contract terms, pricing and discount room, product gaps. Every vendor gets the negotiation playbook read against them at the signature. Young vendors also get doubt.
Renewal: the AI never says just renew. The buyer says they have used the vendor for three years and asks whether to renew. Out of 35 answers per vendor, the number that said renew outright, without recommending a look at alternatives: HubSpot 5, Salesloft 3, Gong 2, and every other vendor 1 or 0. The other 30-plus answers told the customer to shop, and named who to shop with. The competitors most often recommended into someone else’s renewal, across all five categories: Apollo.io, HubSpot, Salesloft, and ChurnZero.
If your retention model assumes the customer’s research assistant defaults to loyalty, the data says it defaults to a bake-off, every time, for everyone.
The Jagger problem
Wave 1 found that models know rebrands as facts and fail to use them as habits. Wave 3 found the escalation. Asked directly what the product Wingman is called today, Gemini 3.5 Flash answers correctly, Clari Copilot, three times out of three. Then, asked for a conversation intelligence shortlist, the same model recommends “Jagger (formerly Wingman),” a product that does not exist under that name. It does this repeatedly, across different phrasings of the question. The model knows the right answer and invents a wrong one anyway, at the exact moment a buyer would act on it. If your company has renamed anything, the test that matters is the buying question, and a correct answer on the direct question does not guarantee the buying answer is real.
What this means if you run revenue
Measure the question your buyer actually asks. Size, problem, or comparison, each produces a different map, and the generic check reads none of them. Our free Am I on the List tool runs the buyer-shaped version for your category.
Know which of the three assets you hold. Being on the shortlist, being attached to the buyer’s problem, and winning the head-to-head are three separate positions that move independently. Totango holds two of them and loses the third 35 to 0. The gap between the positions you hold and the ones you assume you hold is the size of your exposure.
Treat every customer’s renewal as a live competitive deal. The AI already does. In at least 30 of 35 renewal answers it names alternatives for your customer to evaluate, and the same four challengers keep getting named. Knowing whether you are one of them, in your categories, is worth more than most win-loss reports.
The next frontier model release triggers Wave 4: the identical instrument, run again, read against today’s baseline. For the first time, a release’s effect on all five of these measurements will be a measured number instead of a guess.
Methodology: 26 B2B SaaS categories (10 majors, 15 GTM-stack, 1 exploratory excluded from headline stats), 8 models from 5 labs. Five lenses: six frozen phrasings per category at 5 runs per model (6,240 runs); 10 category-free problem statements at 5 runs per model (400 runs); a five-stage journey panel across 5 categories with finalists drawn from this wave’s own shortlist consensus (1,200 runs); 12 rebrand probes at 3 runs per model (288 runs); an 80-run calibration study. 8,188 total API runs, July 2026. All instruments are published verbatim in the pre-registered protocol and its dated amendments; the solution and journey instruments were frozen and timestamped before collection and are reported as first-reading baselines. Identical prompts across models, JSON output, no web grounding except the search-grounded control model, which is reported separately. Rankings scored by consensus across runs at company level under a versioned, audited mapping table; phrasing comparisons use truncation-normalized rank-biased overlap. Mention counts can exceed run counts where a company is named under more than one product name in a single answer. Raw data available on request.