Every week we send the same 20 buyer questions to ChatGPT, Perplexity, Gemini, Grok, and Claude, and record which brands each engine names and which sources it links to. This is the update for the cycle ending August 18, 2026. It is the 20th round since March.
Three items. The engines produced 81 brand citations, the highest total we have recorded, and they did it on 95 of 100 possible engine-question runs rather than a full set. One engine, Gemini, supplied all five of those missing runs, and its absence cut the headline consensus rate in half without a single engine changing its mind. And a correction: a domain alias we registered last cycle was never applied to the rounds collected after it, which changes three figures we published.
The Bottom Line
The number of citation slots available in a week is not fixed. It moved 13% in seven days, and it moved because four of the five engines went up at once. If you track your own citation count in isolation, a flat week here would read as a loss, because the pool you were measured against got bigger. The second half of the point is the one that keeps recurring in this series: any AI visibility number is only interpretable next to how many of the runs behind it actually completed.
Citations Hit 81, From an Incomplete Week
Our totals over the last six rounds read 73, 68, 64, 73, 72, and 81. The previous high was 78, set in June on a complete run of 100. This round produced 81 from 95, which puts citations per completed run at 0.85, also the highest we have recorded.
| Engine | Citations last round | Citations this round | Questions answered | Sources retrieved |
|---|---|---|---|---|
| Grok | 22 | 23 | 20 | 580 |
| Perplexity | 15 | 21 | 20 | 393 |
| ChatGPT | 14 | 21 | 20 | 118 |
| Claude | 7 | 9 | 20 | 122 |
| Gemini | 14 | 7 | 15 | 149 |
Perplexity's 21 is its highest in 20 rounds, breaking a four-round run of 14, 14, 15, 15. ChatGPT's 21 is its best since late July. Grok's 23 ties its own high. Claude's 9 is its best on both citations and distinct brands since early July.
The retrieval column is the part worth sitting with. ChatGPT and Perplexity produced the same number of citations from source pools that differ by a factor of 3.3, and Grok needed 580 sources for a comparable result. ChatGPT links to a brand's own website in 22.9% of its sources, roughly three times the rate of any other engine, which is the mechanical difference between reading a lot and linking to vendors.
One reading is one reading. Three of the last four rounds sat between 64 and 73, so this is a step outside a stable band and not a direction. We will report it as a trend when it survives a third reading, and not before.
Corrections and Revisions
Corrected: the ConvertKit domain alias, applied across the archive
Last cycle we registered kit.com as an alias for ConvertKit, which rebranded to Kit. We applied it to the two historical records that prompted it and did not backfill it into the rounds collected after the rebrand.
Rescanning all 20 rounds with the alias active adds one ChatGPT citation each to three previous weeks, every one of them a help.kit.com page returned for "what email tool should I use for newsletters."
Three published figures change:
| Figure as published | Corrected figure |
|---|---|
| ChatGPT produced 13 citations last cycle, its lowest in 19 rounds | 14 citations, tying the previous low rather than setting a new one |
| Total citations last cycle: 71 | 72 |
| ConvertKit: 13 consecutive rounds at zero citations | Cited in five rounds, including three of the last five |
ConvertKit is back at zero this round, with 8 mentions and no source. The rule we now apply: any change to a brand's identity or to the matching logic triggers a full rescan of every round in the archive, and the resulting differences get published here. This is the second correction of that shape in two cycles, and both came from a rule being registered without being applied uniformly.
Revised: how we report the consensus rate
We have been reporting one consensus number, the share of questions where all five engines name the same brand first. That number is unusable in a week where an engine does not answer, and this was such a week.
| Measure | Last round | This round |
|---|---|---|
| Strict agreement, all five engines answering and agreeing | 30% | 15% |
| Availability adjusted, all answering engines agreeing | 30% | 25% |
| Strong consensus, four or more of the answering engines | 65% | 55% |
Two of the three questions that lost unanimous status were "Vercel vs Netlify comparison" and "best hosting for Next.js apps." On both, all four engines that answered returned Vercel. Nothing changed in what the engines said. Reported on the strict measure alone, this week would read as consensus halving, which would be an infrastructure event described as a market event. From now on both numbers appear here, and the strict series carries a marker on every round where an engine failed.
Also Worth Noting
- Perplexity retrieved 393 sources after 401 last cycle, holding above the 350 threshold we set as the test, and its citation count moved with the pool for the first time. Two readings on both. By our own rule that is a streak, not a trend.
- Gemini answered 15 of 20 questions, all five failures being rate limits, and all four developer tools questions among them. Its answered-question count across the last ten rounds reads 12, 3, 20, 2, 20, 20, 20, 20, 19, 15. No other engine has missed a question since early July.
- Five brands were named by every engine that answered and cited by none: Google Analytics, Monday.com, ConvertKit, Railway, and Fly.io, 48 mentions between them. Two of the five are enterprise brands, which continues to rule out company size as the explanation. Google Analytics and Railway have now gone 20 rounds without a single citation.
- Every one of the 25 tracked brands was named by at least one engine, the first time that has happened in 20 rounds. The two brands that had been invisible for two straight weeks each got exactly one mention, both from Gemini, both inside a longer list with no source attached.
- 7 of 11 startups earned at least one citation, tying the widest startup coverage we have recorded, while the startup share of all citations stayed at 16%, inside its 14 to 19 band. Breadth and share are different questions, and breadth is the one a single challenger brand can move.
- Project management produced 6 citations from 63 mentions. It has had the lowest citation count and the worst mentions to citations ratio of the five categories in six of the last seven rounds. The brand that was the leading answer to two of the four questions earned one citation. The brand that lost both earned three.
- Reddit citations fell to 76 across all engines from last cycle's record 89. Grok supplied 36 and Perplexity 34, the closest the two have ever been. Claude has now gone 20 rounds without citing Reddit once.
What We Are Watching Next
Perplexity. Two readings now show a source pool above 350 and one shows a citation count above its long run band. The condition for next cycle is explicit: a third reading above 350 sources with citations above 16 makes this a platform change we will write about properly. A fall back under 250 sources, or citations returning to the 14 to 16 band, makes it a two week excursion, and we will say so here.
Methodology
20 natural language buyer questions across 5 B2B SaaS categories (CRM, project management, email marketing, analytics, developer tools), covering 25 brands tagged enterprise, midmarket, or startup. Each question goes to ChatGPT, Perplexity, Gemini, Grok, and Claude through live API calls, not cached results, on a weekly cycle since March 2026. This is round 20.
We record which brands each engine names, the order it names them in, and every source URL it returns. A brand counts as "cited" only when a source URL on that brand's own domain appears in the response, and as "mentioned" when it is named without one.
This round returned 95 of 100 engine-question runs. Gemini failed 5 questions after repeated rate limits: all four developer tools questions and one analytics question. Every other engine returned 20 of 20. Where a comparison is affected by those failures, both the raw and the adjusted figure appear above.
Two limitations worth stating. Citation totals are not directly comparable across rounds with different completion rates, which is why we now report citations per completed run alongside the total. And we only observe outputs. We have no visibility into what any engine does internally, so nothing here explains why a number moved.