TL;DR: This AI-first B2B SaaS startup came to us looking for more visibility, with the goal of being recommended alongside industry leaders. When we started working with this brand, the models named the ‘usual’ brands. We built an entire content infrastructure across YouTube, Reddit, and Third-Party editorial. Twelve cycles later the brand had earned 190 AI citations and 91 first-page rankings, and was being recommended in the same answers as their industry-dominant incumbent. Sixteen cycles in, the count is 281 and climbing.
- 0 → 190AI citations in 12 cycles
- 91First-page rankings
- 89Owned assets built
- At parWith the incumbent
The situation
Some brands have a reputation problem. Ghosts have an authority problem.
This client is an AI-first B2B SaaS startup selling into a business-software category that two household names have owned for decades. The product was good and the customers were real, but when a buyer asked ChatGPT, Perplexity or Google AI Overviews “what’s the best software for this job?” the answer listed the incumbents and stopped. The brand was not mentioned, not compared, and not positioned as an alternative. It was absent from the consensus that shapes the buying conversations that follow.
What we found
AI answers are built from the consensus that already exists. If third-party sources don’t mention a brand, citing it feels unnatural to the model, no matter how good the product is. Three things were true for this client:
The incumbents owned the defaults. Decades of coverage, reviews and comparisons had trained the models to surface the same two brands. A startup has none of that history to work with.
There was almost nothing to build on. The initial audit found a minimal footprint outside the brand’s own site: little third-party coverage, next to no presence in AI Overviews, and no consistent appearance in any of the chat assistants. That sounds bad, and it was, but it also meant the first credible third-party assets would not be competing with older, weaker mentions of the brand. They would be the record.
Real buyers’ questions helped us expand efforts. To track our efforts and learn more about how their customers were finding them in search, the client followed up every new customer with a survey during our engagement. When new customers found them via ChatGPT, they asked specifically for the exact prompts and to share the conversation they had to find the recommendation.
Getting real human data like this helped us build a deeper, more personalized content strategy aligning with specific use cases, and ICPs we could target in the content.
Recommendation
Don’t just track how your brand shows up in search results. Get real data from your clients and customers. This can be done via a post-purchase survey, or interview.
- [incumbent] alternatives
- [brand] vs [incumbent]
- best [category] software
- [brand] review
- is [brand] worth it
- [category] for [company stage]
You can’t manufacture authority overnight, and you can’t ask for attention. You build into the conversations buyers are already having.
What we did
For a Ghost, editorial and video carry the first wave of authority. Reddit reinforces it, and it’s where buyers gut-check a shortlist before they sign. We ran the three in parallel from cycle one, at a steady cadence of roughly seven to nine new assets a cycle.
-
Leg 01 · Authority signal
Niche editorial
Long-form comparison, alternatives and best-of articles on an editorial publication we own, with its own readership and backlink history. Targeted at incumbent-adjacent queries, not the brand’s own name.43 articles · 29 page one on Bing
-
Leg 02 · Context signal
YouTube
Head-to-head video reviews against each of the major incumbents, on a channel with posting history, recognised hosts and video-pack authority.38 videos · 27 page one on Google
-
Leg 03 · Consensus signal
Reddit
Community posts, Q&A campaigns and targeted comments where buyers ask the category question. Real creators who use the tools.6 posts · 2 Q&A · 31 mentions
The first cycle-one quick win came from video, where we were able to do full demos and answer specific questions for users from the POV of our industry expert. The first product review we published ranked first in the Google video pack and on page one of Bing within weeks, and was picked up as a featured video in ChatGPT before the second cycle began.
Not every agency can run this play. Building consensus needs owned infrastructure across three channels, an editorial bench able to produce dozens of comparison pieces in parallel, and a pipeline that keeps all three moving at once. We’ve operationalized that over thirteen years of building publishing businesses.
What happened
Citations arrived faster than we expected for a category this locked up, but the growth and ‘halo effect’ took time. Once this brand became a known entity for these specific user queries, we saw more visibility gains.
050100150200123456789101112cycle 4: 24 (down 1)190cycle
- Cycle 1First assets published. The review video ranks first in the Google video pack and appears as a featured video in ChatGPT.
- Cycles 2 to 3Citations begin: 11, then 14 more. Page-one rankings on Bing and Google for the first comparison articles.
- Cycle 4Down one. The models re-evaluate what they have picked up. The month that tests a month-to-month client.
- Cycles 5 to 9Steady climb, 4 to 20 new citations a month, as the comparison library fills out across the incumbents.
- Cycles 10 to 12The inflection: 25, 52 and 33 new citations. The models start cross-referencing the library and each asset accelerates the last. 190 cumulative.
- Cycle 16 (ongoing)281 citations, 119 assets. Pages cited by three or more engines up from 14 to 42.
By cycle twelve the brand had 190 AI citations across ChatGPT, Copilot, Perplexity, Google AI Overviews, Bing AI Overviews and Gemini; 89 owned assets (43 comparison articles, 38 category videos, 6 community posts and 2 Q&A campaigns); and 91 first-page search rankings, 50 on Bing and 41 on Google.
- ChatGPTBest-of and comparison answers
- PerplexityResearch queries and alternatives lookups
- CopilotProduct-comparison responses
- Google AI OverviewsBranded and category queries
- Bing AI OverviewsCategory and comparison results
- Google video packComparison and review queries
The hero moment is easy to describe. Ask the same category question that returned two incumbents at the start, and the answer now names the challenger in the same breath as the dominant one.
Before · cycle 0
“What’s the best software in this category?”
- Incumbent A
- Incumbent B
- The challenger
After · cycle 12
“What’s the best software in this category?”
- Incumbent A
- The challenger
- Incumbent B
A brand with a fraction of the incumbent’s spend, cited by AI as a peer.
The compounding effect. Gemini was the slowest surface to pick the brand up: four citations at cycle twelve against dozens on Perplexity and Copilot. It has since caught up (nineteen by cycle sixteen), but anyone planning a program like this should expect the engines to move at different speeds, and should not judge the whole effort by the slowest one. This is why we track across the LLMs. There is often a compounding ‘halo effect’ from these efforts that builds over time.
Where it stands now
The engagement is ongoing. Sixteen cycles in, the library is 119 assets and the citation count is 281. The number of pages cited by three or more engines at once, the clearest sign that a source has become part of the consensus rather than a one-off, has tripled since cycle twelve, from 14 to 42.
On the independent tracking we run alongside our own, the brand now holds a 24% share of voice across the category’s buyer prompts, within two points of the number-two incumbent and ahead of it on average position in the answer, with sentiment on a par with both leaders. Third in a category where it was nowhere.
What we learned
Ghost problems are authority problems. AI cites what the consensus already says. To change the answer you change the source material the models learn from, and that means third-party content.
Comparison is the fastest route in. Buyers ask AI to compare incumbents and alternatives. Own those lanes and the model surfaces you alongside the leaders.
Don’t forget about your real customers. Being able to see what real users looked up in ChatGPT helped our future content efforts. We recommend that brands start to ask more specific questions in post-purchase surveys. Even better if you can actually see their search logs.
Early cycles build the library. Later cycles cash it in.
What’s next
Maintain search presence. The next cycles extend the comparison library into the company-size and use-case questions buyers ask once they have a shortlist, and keep the video channel current as the incumbents ship new features. The position is earned; keeping it is the work.
We’ve also started a regular optimization and updating cadence to maintain and strengthen the current positions and brand mentions in search. Keeping on top of new product features, use cases and new competitors also gives us more room to grow and expand.
Is your brand a Ghost?
Run your category’s best-of question through ChatGPT, Perplexity and Google. If the answer names your competitors and not you, that is the problem this playbook solves. Get a proposal and we’ll run the full prompt set for you.
How we measured this
Citations are tracked two ways. Our internal tracker takes a monthly snapshot of the category’s core buyer journeys (comparison, alternatives, best-of, review and fit-for-stage queries) across Google, Bing, ChatGPT, Perplexity and Copilot, with Google AI Overviews, Bing AI Overviews and Gemini added as those surfaces matured. A citation is counted once per asset per engine when the brand, or an asset we built for it, is referenced in the answer.
Every asset, including older ones, is re-checked from scratch each cycle, so the numbers reflect what the engines are citing now, not what they cited once. Search rankings are first-page positions on Google and Bing for the target queries. Share of voice, average position and sentiment come from Peec.ai, an independent AI visibility platform, over the same category prompt set.
We only count one citation per source per asset in our reporting, but every asset contributes to many more citation sources, so the effort is compounded and expanded.
Frequently asked questions
Why is the client anonymous?
We anonymize case studies by default so clients can share results without disclosing competitive strategy. The category, stage, method and numbers are real; the name is withheld.
Does this work for any B2B SaaS brand?
It works for brands with a good product in a category where buyers compare options and ask AI for recommendations. The play depends on producing credible third-party comparison content, which needs a product that holds up under honest review.
How long before AI citations start?
In this engagement the brand first appeared in an AI answer in cycle one, citations were being counted from cycle two, and the count turned sharply upward around cycle ten. Timing varies with category competitiveness and how quickly assets are indexed, so we don’t promise a cycle count.
Does a brand’s own blog count?
It matters for other reasons, but it’s not what moves AI answers (at least on its own). Models weight independent third-party sources when forming a consensus, which is why the work happens on editorial publications, video channels and communities rather than on the brand’s site.
What did the engagement cost?
This one ran on a 35-hour monthly retainer, month to month. Scope depends on the category and the number of incumbents to compare against; we quote per engagement.