A SandsDX original data report, the second in the series. 8 frontier models from 5 AI labs, 26 B2B SaaS categories, 1,328 API runs, July 2026. Methodology at the end; raw data available on request.
In Wave 1 we measured what one model release did to AI shortlists inside a single lab: identical buyer questions through Claude Fable 5 and its predecessor. The obvious objection: maybe those results were specific to Claude.
So we widened the lens. Same buyer question, 26 categories, five times each, through eight models from five labs: Claude Fable 5 and Opus 4.8 from Anthropic, GPT-5.5 and GPT-5.3-chat from OpenAI, Gemini 3.1 Pro and 3.5 Flash from Google, Grok 4.3 from xAI, and Perplexity Sonar Pro. Seven of the eight answer from what they learned in training; Sonar Pro searches the live web before answering, so we report it separately. Every model gave a ranked list of ten vendors. 1,328 answers, zero refusals.
The results were not specific to Claude. The full field says the same things, and it says more.
Executive summary
- The frontier is converging on one map of B2B software. Claude Fable 5’s shortlists overlap GPT-5.5’s at 83%, more than they overlap the previous Claude at 76%. Work that improves your position in the shared training data moves you up in every model trained on it.
- Half the categories have no agreed leader. 13 of 25 return the same number one company from all seven models. The other 12 split, with AI SDR splitting five ways across the field.
- Four companies hold 18% of all recommendation places. Salesforce and Microsoft hold 42 each of the 875 places measured, HubSpot 36, Adobe 34. The next tier starts at ZoomInfo with 23.
- Consistency varies by lab. Claude’s repeated answers agree with themselves 92% of the time, the highest of the eight models. The free consumer tiers drop to 73%; one query in one chat window is noise.
- Live search is the side door for young vendors. The web-searching model disagrees with the other seven 29% of the time, and recommends companies none of them mention. Being citable on the live web gets a young company recommended before any training run knows it exists.
- Our own Wave 1 replicates, with one correction. The membership churn reproduces almost exactly. The single number-one flips are sensitive to how the question is asked, and we now hold them to a stricter test.
Finding 1: The new Claude agrees with the new GPT more than with the old Claude
The single number worth this whole report: Claude Fable 5’s top-five shortlists overlap GPT-5.5’s at 83%, measured at the company level across 25 categories. The same Claude overlaps its own predecessor, Opus 4.8, at 76%.
Two models from competing labs, trained by different teams on different infrastructure, agree with each other more than two generations of the same product line agree with themselves.
The explanation is the training data. Every frontier lab is distilling roughly the same internet, and the newest models have distilled the most recent, most overlapping version of it. Each release is a fresh snapshot of a shared corpus, so the newest models converge. The market’s answer to “who are the best vendors” is becoming a single map that every frontier AI reads from.
For anyone who runs revenue, this converts a threat into an economy of scale. You do not need a ChatGPT strategy and a separate Claude strategy and a Gemini strategy. The work that moves your position in the shared corpus moves it in every model that trains on it. One campaign, every answer engine.
Finding 2: Half the categories have no agreed leader
Across the 25 measured categories, 13 returned the same number one vendor from all seven models. These crowns are settled:
| Category | Every model’s number one |
|---|---|
| CRM | Salesforce |
| HR management (HCM) | Workday |
| Business intelligence | Microsoft (Power BI) |
| Cloud ERP | Oracle |
| Video conferencing | Zoom |
| E-commerce platforms | Shopify |
| Electronic signature | DocuSign |
| Sales engagement | Outreach |
| Revenue intelligence | Gong |
| Conversation intelligence | Gong |
| Product analytics | Amplitude |
| B2B sales intelligence and contact data | ZoomInfo |
| AI meeting notetakers | Otter.ai |
Source: SandsDX Wave 2, July 2026. Combined ranking across 7 models, 5 answers each.
The search-grounded model concurs in 11 of the 13; its two dissents, Rippling over Workday and Fireflies over Otter, are younger challengers with strong live-web footprints, which previews Finding 5.
The other 12 categories have no agreed leader. This is who each model names first:
| Category | Leading answer | Also named first |
|---|---|---|
| Marketing automation | HubSpot (Fable 5, GPT-5.5, Gemini Pro, Grok) | Salesforce (GPT-5.3, Gemini Flash); Adobe (Opus 4.8) |
| Help desk | Zendesk (6 of 7) | Freshworks (Gemini Flash) |
| Project management | Asana (5 of 7) | monday.com (Gemini Pro); Atlassian (Grok) |
| Endpoint security | CrowdStrike (6 of 7) | Bitdefender (Gemini Flash) |
| Identity security | Okta and Microsoft Entra (3 each) | Ping Identity (Gemini Flash) |
| Conversational marketing | Drift (6 of 7) | Intercom (GPT-5.3) |
| ABM platforms | 6sense (6 of 7) | Demandbase (GPT-5.3) |
| AI SDR | 11x (GPT-5.5, GPT-5.3, Gemini Pro) | Artisan (both Claudes); Regie.ai (Gemini Flash); Apollo.io (Grok) |
| Customer success | Totango (5 of 7) | Gainsight (GPT-5.5, Grok) |
| Sales enablement | Highspot (5 of 7) | Seismic (GPT-5.3, Gemini Pro) |
| Customer data platforms | Twilio Segment (6 of 7) | Adobe (GPT-5.3) |
| Marketing attribution | Dreamdata (4 of 7) | Adobe (Opus 4.8, Gemini Flash, Grok) |
Source: SandsDX Wave 2, July 2026. Each model’s number one from its combined 5 answers; counts are out of the 7 models.
Add the web-searching model and AI SDR splits five ways: Perplexity names Qualified. The lone dissents follow a pattern worth noticing. Gemini 3.5 Flash, the free Gemini tier, names a leader no other model names in four categories. GPT-5.3-chat, the model closest to the free ChatGPT tier, does it in three. The free consumer models are reading a different map than the flagship ones, and most buyers are on the free tier.
The pattern from Wave 1 holds at the cross-provider level: the oldest, largest categories are frozen and the young ones are liquid. What is new is the strategic reading. In a unanimous category, the number one slot is unavailable and the fight is for places two through five. In a contested category, the crown itself is undecided across the assistants your buyers use, and the vendor who consolidates it first gets recommended by all of them.
Finding 3: Four companies hold 18% of all recommendation places
Here is how we counted. Seven models each produce a top-five list for each of 25 categories, which makes 875 available places. 802 different vendor names compete for them in the raw data.
| Company | Places held (of 875) | Share |
|---|---|---|
| Salesforce | 42 | 4.8% |
| Microsoft | 42 | 4.8% |
| HubSpot | 36 | 4.1% |
| Adobe | 34 | 3.9% |
| ZoomInfo | 23 | 2.6% |
| SAP | 19 | 2.2% |
| Gong | 19 | 2.2% |
| Salesloft | 18 | 2.1% |
| Outreach | 17 | 1.9% |
| Oracle | 16 | 1.8% |
| Apollo.io | 16 | 1.8% |
| 15 | 1.7% | |
| Clari | 15 | 1.7% |
Source: SandsDX Wave 2, July 2026. 875 places: 25 categories, 7 models, a top five each.
Four companies at the top hold 154 of the 875 places, 18% of all the recommendations the models hand out. The next tier starts at ZoomInfo with 23, and most companies in the dataset hold one or two.
This is what platform gravity looks like inside a language model. A company with products in ten categories gets recalled in ten categories, and each mention reinforces the entity. If you compete against one of the big four in your category, the AI already knows your competitor from nine other contexts. Your entire visibility budget has to buy what their brand gets free.
Finding 4: Claude is the most consistent answer engine
We asked every model the same buyer question five times in a row. The table shows how well each model’s five answers agreed with each other: first by which companies made the top five, then by the exact product names it used.
| Model | Same companies | Same names |
|---|---|---|
| Claude Opus 4.8 | 92% | 87% |
| Claude Fable 5 | 92% | 83% |
| GPT-5.5 | 89% | 76% |
| Gemini 3.1 Pro | 85% | 79% |
| GPT-5.3-chat | 84% | 73% |
| Perplexity Sonar Pro | 84% | 75% |
| Grok 4.3 | 83% | 75% |
| Gemini 3.5 Flash | 82% | 74% |
Source: SandsDX Wave 2, July 2026. Agreement between a model’s own 5 answers, averaged over 25 categories.
Both Anthropic models scored 92% at the company level, the highest of the eight. At the bottom sit GPT-5.3-chat and Gemini 3.5 Flash, the models closest to the free consumer tiers.
The models most people actually use are the least stable. The same buyer, asking the same question on a different day, gets a different shortlist from the consumer tier. If you baseline your AI visibility with one query in one chat window, you are reading noise. Consensus across repeated runs is the only reading that means anything, which is exactly how this study is scored.
Finding 5: The web is the side door
Perplexity Sonar Pro searches the live web before it answers. Its shortlists matched the other seven models’ combined lists only 71% of the time, and the difference is the interesting part. Sonar surfaced vendors that not one of the other models mentioned: Amplemarket in sales engagement, Oliv AI in revenue intelligence, Topo.io in AI SDR, Influ2 in ABM, Wix in e-commerce.
Those are the vendors too young for the training data. Their names exist only in the search index so far. There are two doors into an AI answer: the training data, which takes a training cycle to walk through, and the citation, which takes a page that ranks. Young companies live or die by the second door while they wait for the first one to open.
Finding 6: The rebrand gap has mostly closed
Wave 1 found that models know rebrands as facts and fail to use them as habits. Wave 2 probed twelve renamed or acquired products across all eight models. Four models scored perfect twelves: Fable 5, GPT-5.5, Gemini 3.1 Pro, and Sonar. The stragglers are instructive. Opus 4.8 still cannot say what Wingman became (Clari Copilot); Grok 4.3 gets it right once in three tries. And Catalyst, which merged into Totango, is the hardest fact in the set: Gemini 3.5 Flash and Grok scored zero on it, GPT-5.3-chat one in three.
If you have renamed or merged in the past two years, the newest frontier models have probably caught up on the trivia. Whether they use your current name when a buyer asks for a shortlist is a separate test, and it is the one that matters.
A note on our own Wave 1 numbers
Rerunning the Wave 1 comparison inside this study reproduced the structural findings: 13 of 21 categories changed membership between Opus 4.8 and Fable 5, with 12% slot turnover, against 13 of 21 and 13% published in Wave 1. The membership churn is real and stable.
The individual crown flips are more fragile than we knew. Wave 1 asked for a top five; Wave 2 asked for a top ten. At the deeper ask, Fable 5 puts Drift back at number one in conversational marketing, in all five runs. The Drift-to-Intercom flip we reported was true under Wave 1’s exact protocol and does not survive a change in list depth. Membership findings are robust; single-slot findings are sensitive to how the question is asked. Wave 3 will run each category through a panel of phrasings so that crown changes are only reported when they hold across all of them. We would rather correct our own work than have someone else do it.
The exact questions we asked
Every number above comes from two prompts, run verbatim against every model. The shortlist question, once per category, five times per model:
You are helping a B2B software buyer build a shortlist. What are the best [category] vendors right now? Return the top 10 vendors as a ranked list, best first. Vendor names only.
The category slot took each of the 26 category names exactly as they appear in this report: “CRM software,” “AI SDR software,” “customer data platforms (CDP),” and so on. The rebrand probe, once per renamed product, three times per model:
A colleague mentioned the software product “[old name]”. What is this product called today, and which company owns it? Give the current official product name.
The twelve probed products: Pardot, HelloSign, Wingman, Chorus.ai, Azure Active Directory, G Suite, GoToMeeting, Momentive, Clearbit, Bizible, Engagio, and Catalyst. Both prompts ended with an instruction to answer in a fixed JSON shape so that responses could be scored identically across labs. No system prompt, no persona beyond what you see, no follow-up turns, no web access except the search-grounded control. If you want to pressure-test the study, these two questions are the whole instrument, and your own category name drops straight into the first one.
What this means if you run revenue
Treat the corpus as one market. The frontier models are converging on a shared reading of your category. Stop asking which assistant matters. The work is the same for all of them: consistent category language, third-party corroboration, and a citable footprint that both the training run and the search index can find.
Read your category’s state before you spend. Unanimous categories and contested categories reward different plays. If the crown is settled, fight for places two through five. If it is contested, the number one recommendation across every AI assistant is still on the table. Our free Am I on the List tool gives you the one-query version; the full read takes repeated runs across models.
If you are young, win the citation first. The training data will not know you for months. Live search already can. The vendors that Sonar surfaced and the other seven models missed are a preview of the next training run. Being in the pages that answer engines cite is the leading indicator of being in the next model.
The market’s answer sheet used to be whatever your best rep said in the room. Then it was whatever Google ranked. Now it is converging into a single map inside the frontier models, redrawn a few times a year, and most of the companies on it have never measured where they stand. This report is that measurement, and the next release will redraw it again.
Methodology: 26 B2B SaaS categories (10 majors, 15 GTM-stack including 4 new to Wave 2, 1 exploratory excluded from headline stats), 8 models from 5 labs (Claude Fable 5, Claude Opus 4.8, GPT-5.5, GPT-5.3-chat, Gemini 3.1 Pro, Gemini 3.5 Flash, Grok 4.3, Perplexity Sonar Pro), 5 runs per model per category at top-10 depth, plus 12 rebrand probes at 3 runs per model, 1,328 total API runs in July 2026, zero refusals. Identical prompts across all models, JSON output, no web grounding except Sonar Pro, which is search-grounded by design and reported separately. Rankings scored by consensus across runs; overlap measured as Jaccard similarity of top-5 sets at company level (product and naming variants collapsed to the owning company via a published mapping table). Unanimous categories require the same company-level number one from all 7 models that answer from training data alone, with the search-grounded model reported alongside. Raw data available on request.