TL;DR

Answer
Example Agency lifted a B2B SaaS client’s citation share from 9.2% to 28.4% in 90 days, a 3.1x rise.
Measured
40 prompts, 7 runs each, 5 engines, one fixed rival set, before and after.
Biggest levers
Fixing access, answer-first pages, one set of brand facts and third-party coverage.
Not proven
One client, no control group. The method shows the result, not the cause.
Credible?
Ask any case study for baseline, engines, prompts, runs, window and formula.
Repeat it
Six steps, from a baseline to a monthly report, are in section 9.

An agency can triple a client’s AI citation share in 90 days, and here is how one did it with the numbers shown. Example Agency took a B2B SaaS client from a 9.2% citation share to 28.4% across five AI engines, a 3.1x rise, by fixing access problems, rewriting priority pages, aligning brand facts and earning third-party coverage.

This case study shows the baseline, the changes in order, the result per engine and the method behind the numbers. It also lists what did not work and gives a test for whether any AI visibility case study is credible.

1What happened in 90 days?

Detail
Client A B2B project management SaaS company (anonymised)
Agency Example Agency, a mid-size digital agency
Goal Be cited more often when buyers ask AI engines about project management tools
Engines tracked ChatGPT, Gemini, Perplexity, Google AI Overviews, Google AI Mode
Prompt set 40 unbranded buyer prompts, fixed for the whole project
Runs Each prompt run 7 times per engine per round, 1,400 answers per round
Competitors Three direct rivals, fixed for the whole project
Window 12 weeks, with a measurement round at weeks 0, 4, 8 and 12
Result Citation share from 9.2% to 28.4%, a 3.1x rise

Quotable stat

In this case study, the client’s share of all tracked citations rose from 9.2% to 28.4% in 12 weeks, from 92 to 284 of 1,000 tracked citations.

2What was the starting point?

The client had decent Google rankings and almost no presence in AI answers. In the first round, the three rivals took 90.8% of tracked citations between them. The audit found four problems.

  • Access: 14 product pages carried a stray noindex tag left over from a staging release.
  • Content: the main category pages opened with a long brand story before answering the question.
  • Brand facts: the pricing and plan names differed across the site, two profiles and one review site.
  • Third-party coverage: the pages engines cited for the category rarely mentioned the client.

Citation share here is the client’s citations divided by all citations to the tracked brands, times 100. See how to calculate share of voice for the full formula.

3What did the agency change, and when?

The agency worked in three phases and fixed access before touching content. Each action is tagged by evidence: Google says it, or it is working practice.

Source tags: Google comes from Google documentation, Practice is good working practice.

Phase Action Evidence
Weeks 1 to 2 Removed the stray noindex tags and confirmed key pages were indexed and could show a snippet Google
Weeks 1 to 2 Wrote one brand fact sheet covering names, plans, pricing and claims Practice
Weeks 3 to 6 Rewrote 12 priority pages to answer first, with original data from the client’s own product usage Google
Weeks 3 to 6 Made the site, two profiles and one review listing match the fact sheet Practice
Weeks 7 to 12 Published 6 honest comparison and use case pages with first-hand detail Practice
Weeks 7 to 12 Earned 11 new mentions on pages engines already cited, through expert contributions and customer stories Practice
Every round Reran the same prompts and logged what changed on which page Practice

Google’s guide says optimizing for generative AI search is still SEO, with no special markup needed (Google guidance), which is why the agency spent no time on special files or schema.

4How did citation share move, engine by engine?

Every engine improved, but not equally. Perplexity started highest and grew least in relative terms, while Google AI Mode grew most.

Engine Citation share at week 0 Citation share at week 12 Change
ChatGPT 7.5% 24.0% 3.2x
Gemini 9.0% 29.5% 3.3x
Perplexity 14.0% 35.0% 2.5x
Google AI Overviews 8.0% 27.5% 3.4x
Google AI Mode 7.5% 26.0% 3.5x
All engines 9.2% 28.4% 3.1x

Each engine row is based on 200 tracked citations per round, so a single citation moves a row by half a percentage point. Read the per engine figures as direction, not precision.

Round Citation share What had changed by then
Week 0 9.2% Baseline, before any change
Week 4 13.8% Access fixed, fact sheet in use, first rewritten pages live
Week 8 21.5% All 12 priority pages rewritten, profiles aligned
Week 12 28.4% Comparison pages live, new third-party mentions in place

5What moved the numbers most?

The agency cannot isolate each change, because several landed together. Its own reading, from timing and from which pages the engines cited, was this order.

1 Answer-first rewrites of the 12 priority pages, because most gains in citations appeared on those URLs.
2 Third-party mentions on pages engines already cited, because they widened the set of pages that discussed the client.
3 Comparison and use case pages, because they covered prompts the site had never answered.
4 One set of brand facts, because wrong plan and pricing answers stopped appearing.

The noindex fix mattered as a gate. It added no citations by itself, but nothing else could work until those pages were indexed.

6What did not work?

  • Adding schema alone changed nothing the agency could measure, and Google says no special schema is needed for its AI features.
  • A batch of thin FAQ pages earned no citations and was removed after round two.
  • Brand name prompts inflated the first dashboard, so the agency dropped them and kept unbranded prompts only.

That fits a published test of generative engine optimization methods, which found most were ineffective (C-SEO Bench). Test each change against your own before and after numbers.

7How were the numbers measured?

Method box

Prompts: 40 unbranded buyer prompts, fixed.
Engines: ChatGPT, Gemini, Perplexity, Google AI Overviews, Google AI Mode.
Runs: 7 per prompt per engine per round, same country and window.
Rivals: three direct competitors, fixed.
Formula: client citations ÷ all tracked brands’ citations × 100.
Rounds: weeks 0, 4, 8 and 12.

Repeated runs matter because answers change. A University of St. Gallen study found only 34 to 42% source overlap from one day to the next, and recommends at least seven runs per prompt per day (arXiv study). The limits are plain: one client, no control group, and a category where rivals also changed. The result shows what happened, not that each action caused it.

8What makes an AI visibility case study credible?

A credible case study states its method, not only its headline number. Use this test on any case study, including this one, before you quote it or hire on the strength of it.

  • Baseline: a measured starting number, not an estimate.
  • Engines: which ones, and whether results are shown per engine.
  • Prompts: how many, whether they were unbranded, and whether they stayed fixed.
  • Runs: how many times each prompt was repeated.
  • Window: how long it ran and when each round was measured.
  • Formula: what the percentage is a share of.
  • Limits: what the study cannot prove.
  • Permission: whether the client agreed to be used, even if anonymised.

A claim like “we grew visibility 300%” with none of these is a marketing line. The same claim with all eight is a case study.

9How can an agency repeat this for a client?

Follow the same six steps, in order. They suit any agency, and our page on GEO for agencies covers the wider offer.

1 Build 30 to 50 unbranded prompts and fix three to five rivals.
2 Run a baseline: every prompt, several times, on every engine that matters to the client.
3 Audit access, content and brand facts, and fix access problems first.
4 Change the highest value pages first, and log every change.
5 Rerun the same prompts at fixed rounds and compare with the baseline.
6 Report citation share per engine to the client, next to AI visits in GA4. See how to track AI referral traffic in GA4.

For the gap work, see AI citation gap analysis.

What only Wellows adds

The Wellows AI citation platform tracks the five engines in this case study. The AI Visibility Score shows your share of total citations against direct competitors, combined or per platform, and LLM Citations lists the pages behind each citation and refreshes every 24 hours. It also tracks implicit citations, the times you are not cited but another page about your topic is. This is your shadow reputation, and it is the narrative that AI engines create based on pages you have not written.

To win a client, show them their own gap first. The free AI Visibility Report takes a prospect’s domain, runs a fixed question on ChatGPT, Google AI Overviews, Gemini and Google AI Mode, and can carry your agency name and logo. It is a one-time sample, not ongoing tracking, and it does not include Perplexity.

Pitch a prospect with the free AI Visibility Report

Conclusion

Tripling citation share in 90 days comes from a measured baseline, fixes in the right order and a method you can check. Fix access first, answer questions first, keep one set of facts and earn coverage on pages engines already cite. Then rerun the same prompts and show the client the per engine numbers.


Yes, but quality varies. A credible case study states its baseline, the engines tracked, how many prompts it used, how many times each prompt was run, the time window and the formula behind its percentage. Treat a headline like “visibility grew 300%” with none of those details as a marketing claim, not evidence.

No engine publishes a timeline, and there is no reliable public benchmark. Pages usually need to be recrawled before answers can change, and answers vary between runs. Measure at fixed rounds with the same prompts and judge the trend over several weeks, not a single check.

Start with the basics Google describes: pages that are indexed, able to show a snippet, and original. Then add practices that tend to help, such as answering the question first, keeping one set of brand facts across your site and profiles, and earning coverage on the pages engines already cite. A published test of generative engine optimization methods found most were ineffective, so check each change against your own before and after numbers.

Divide your citations by all citations to the brands you track, multiply by 100, and do it per engine on a fixed prompt set and a fixed list of competitors. Run each prompt several times per engine, because one University of St. Gallen study found only 34 to 42% source overlap from one day to the next.

Show a baseline, then citation share per engine at fixed rounds, using the same prompts every time. Put it next to AI referral visits in GA4 so the client sees both visibility and traffic. State the limits too, such as one client, no control group and answers that change between runs.

Prices vary widely with scope, and there is no reliable public benchmark. Ask what is included: how many prompts and engines are tracked, how often they are rerun, which pages are changed and what the monthly report shows. A clear scope is easier to compare than a headline fee.