- In our US B2B SaaS case, answers naming the brand went from 128 to 175 out of 1,000 in 44 days, mostly on Perplexity (21 to 50).
- ChatGPT did not move (17 to 17), and two objectives set in advance were missed: both are published with their figures.
- In a SparkToro study, ChatGPT and Google’s AI almost never returned the same list of brands twice across 100 runs, so a single before/after screenshot proves nothing.
- A useful case states the question set, engines, runs per question, dates, the content published and a noise threshold set before comparing.
What should a GEO case study with measured B2B results show?
A fixed set of buying questions, the same protocol run before and after, the number of answers behind each figure, and the change engine by engine, including the figures that did not move. Without those elements, a case study is a testimonial.
The demand for proof is fair. In G2’s 2026 survey, 51% of B2B software buyers said they start research with an AI assistant more often than with Google [1], so AI visibility budgets now get the scrutiny paid search gets. A case you can rely on states six things:
- The question set. How many buying questions, fixed before the first measurement, and which ones were targeted.
- The protocol. Engines, runs per question, dates, and the number of answers per measurement.
- Two measurements on the same protocol. Before and after, with the gap in days.
- What was published. The content produced, so a change can be linked to work.
- The misses. Objectives set in advance and reported even when missed.
- A noise threshold. The variation expected with no work at all, set before any comparison.
What moved in our US B2B SaaS case?
Answers naming the brand went from 128 to 175 out of 1,000 in 44 days. Most of that gain came from Perplexity, and the wider citation count rose mainly on the questions the program targeted.
The client is a US B2B SaaS vendor (ITSM). Both measurements, on August 9, 2026 and September 22, 2026, used the same 50 buying questions, 5 runs each, for 1,000 answers + 250 Google searches per measurement.
| Engine | Answers naming the brand, August 9, 2026 | Answers naming the brand, September 22, 2026 |
|---|---|---|
| Perplexity | 21 of 250 | 50 of 250 |
| Claude | 33 of 250 | 43 of 250 |
| Gemini | 57 of 250 | 65 of 250 |
| ChatGPT | 17 of 250 | 17 of 250 |
| Total | 128 of 1,000 | 175 of 1,000 |
A second, wider count tells the same story. On the 13 questions the program targeted (260 answers per measurement), answers that cited the brand, either by name or as a linked source, went from 11 to 63. The table counts only answers naming the brand. Between the measurements, 3 pieces of content were published, and 54 distinct answers in the second measurement cited at least one of them as a source.
What did not move, and why publish it?
ChatGPT stayed at 17 answers naming the brand out of 250, and two objectives set before the program were missed. Publishing the misses is what lets you trust the gains.
- ChatGPT. The share of answers naming the brand stayed at 6.8%, against a target of 10%.
- Overall citation rate. Answers naming or citing the brand rose from 23.8% to 29%, short of the 30% target.
- The video. Of the 3 published pieces, the video was cited in 0 answers.
- Questions outside the plan. On 8 identity and access questions the program did not target, answers citing the brand went from 66 to 59 out of 160.
The ChatGPT result is the most useful lesson. ChatGPT search retrieves results from third-party providers, and OpenAI says placement is not guaranteed [2]. Gains on Perplexity did not carry over to ChatGPT in the same 44-day window, which is why we report each engine separately.
How do you tell a real result from normal variation in AI answers?
Ask each question several times on each engine, and set, before any comparison, the variation you would expect with no change at all. A single before/after screenshot cannot separate a result from chance.
In a SparkToro study, ChatGPT and Google’s AI almost never returned the same list of brands twice across 100 runs [3]. In the case above, each question ran 5 times per engine, and the noise threshold of ±15% was set on August 9, 2026 by running the full protocol twice the same night, before any comparison. A change inside that band is not reported as a result. In another provider’s case, ask how many answers each figure rests on: “mentions doubled” can mean two answers becoming four.
Which sources did the engines cite, and where did the client’s own site rank?
At the first measurement, youtube.com and reddit.com were the two most-cited domains in the client’s market, ahead of the client’s own site. By the second, the client’s own site had moved from rank 5 to rank 2 among all cited domains.
At the start, youtube.com drew 390 citations and reddit.com 373, across all 5 engines. Vendor sites carry a small share: in one study of B2B software recommendations, the recommended tool’s own site was cited only 11.6% of the time [4]. In Google AI Overviews (250 searches per measurement), the client’s site went from outside the top 10 cited domains to 6th.
Our market-wide measurement shows the same pattern. On November 8, 2026, we asked 65 US buyer questions about GEO, this one included, on ChatGPT, Claude, Gemini and Perplexity, 5 runs each, for 1,300 answers + 325 Google searches [6]. reddit.com was cited on 59 of 65 questions (251 citations), youtube.com on 53 of 65 (235), en.wikipedia.org on 36 of 65, linkedin.com on 34 of 65 and searchengineland.com on 26 of 65. Apply the checklist below to any case an engine shows you.
How should you read a GEO case study from any provider?
Check the base behind each figure, the protocol and the dates, look for published misses, then ask for the raw answers. A provider that will not hand over the answers behind its numbers is asking you to trust a summary.
- “Mentions up 300%.” Ask out of how many answers, on which engines, with how many runs per question.
- “Now recommended by ChatGPT.” Ask for the share of runs that name the brand, not one screenshot.
- “AI traffic doubled.” Ask for the base in visits and the period.
- An anonymous client. Acceptable if sector, market and protocol are stated. Ours is anonymized too.
- No misses at all. Ask which objectives were set before the work, and how each ended.
In the US, the FTC rule on consumer reviews and testimonials bans fake reviews and testimonials [5]; hold case studies to the same standard. Anonymized is fine; unverifiable is not.
How does ZivRank measure a program?
We measure 50 buying questions on 5 engines (ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews), 5 runs per question, before any work and again with the same protocol, and our team writes, fact-checks and publishes the content in between.
Each measurement covers 1,000 answers + 250 Google searches. Changes are reported only beyond a noise threshold of ±15%, and you receive the raw answers with a monthly report, missed objectives included. Programs start at $2,300 per month over 6 months, with a $2,000 baseline audit included in the program and billed at the start. We commit to the work and the method, not to a ranking. The free Quick Scan covers up to 5 of your buying questions, with a video walkthrough by an expert within 48 hours. Read the full case or our measurement methodology.
Frequently asked questions
What methodology should a GEO agency use to measure results?
Agree on the buying questions before the first run, then keep them fixed. Ask each one several times per engine, count named, cited and absent separately, and set aside questions that already contain your brand. Measure the noise by running the protocol twice before comparing anything. At ZivRank, each question runs 5 times and changes within ±15% are not reported as results.
Why can one engine improve while another stays flat?
Each engine retrieves from its own sources and ranks them its own way. A page or mention that Perplexity picks up quickly may not appear in the results ChatGPT search retrieves for the same question. Report every engine separately, and plan specific work for the engine that lags, such as presence on the sources it cites.
Can a GEO agency guarantee results in ChatGPT?
No. OpenAI says placement in ChatGPT search is not guaranteed, and answers vary between runs. What an agency can commit to is the work: questions measured, content published, the protocol used for each re-measurement, and objectives reported honestly, including the ones it missed. A guarantee of rankings is a reason to walk away.
Is an anonymized GEO case study worth anything?
Yes, if it stays checkable. The sector, market, dates, question count, engines, runs per question and the content published should be stated, and the provider should be able to show raw answers on request. Anonymity protects the client; it should never hide the method or the base behind each percentage.
Sources
Every figure in this article comes from the sources below, checked on November 18, 2026. Bracketed numbers in the text refer to this list.
- [1] G2 (Tim Sanders), In the Answer Economy, Don’t Win the Click — Win the Answer, April 15, 2026.
- [2] OpenAI Help Center, Searching the web with ChatGPT, checked October 11, 2026.
- [3] SparkToro (Rand Fishkin), NEW Research: AIs are highly inconsistent when recommending brands or products, January 27, 2026.
- [4] DerivateX, B2B SaaS AI citation study, May 31, 2026, updated September 22, 2026.
- [5] Federal Trade Commission, Trade Regulation Rule on the Use of Consumer Reviews and Testimonials, posted August 21, 2024.
- [6] ZivRank, US measurement: 65 buyer questions about GEO, ChatGPT, Claude, Gemini and Perplexity × 5 runs (1,300 answers) + 325 Google searches, November 8, 2026.