Case Study: From Zero AI Visibility to Cited in 60 Days

Going from zero citations to consistent presence in AI answers takes about 60 days for a mid-sized site, and the work is more structural than promotional. This walkthrough is a representative, anonymized composite of the method rather than one named account: the timeline and actions are the ones that reliably move citation share, and the figures are directional targets grounded in published GEO research, not guarantees. The core lesson holds across every version of it — you win citations by making your best answers extractable, then by making the open web describe you consistently.

The starting condition was typical. A B2B software company in a competitive category had solid classic SEO, ranked on page one for its head terms, and was completely absent from ChatGPT, Perplexity, and Google AI Mode answers to the same questions. Ranking on Google and being cited by an answer engine turned out to be nearly independent problems.

Week 0: baseline and diagnosis

Before touching anything, we recorded a baseline across a cluster of twenty buyer prompts, phrased the way real users ask them. For each prompt we logged which brands the engines named, which sources they cited, and where our subject appeared (nowhere, initially). This baseline is the entire experiment's control; without it, every later claim of improvement is unfalsifiable.

The diagnosis surfaced three failures. The site buried its answers past 400 words of preamble, so retrieval models had no clean passage to grab. Its About page never stated the category, founding year, or headquarters, so engines could not disambiguate the brand. And the open web described the company inconsistently, with three different one-line descriptions across its own profiles.

Weeks 1 to 2: restructure for extractability

The first build phase targeted on-page structure. We rewrote the top ten pages so that each opened with a self-contained answer: the first 60 to 90 words after every question-shaped H2 stood alone as a citable passage. The GEO paper by Aggarwal and colleagues (KDD 2024) measured a 30 to 40 percent visibility lift from adding citations, quotations, and statistics, so we also added concrete numbers and named sources to previously vague sections.

We shipped FAQPage schema on the pages where it fit, matching the rendered HTML to the JSON-LD exactly. Structured Q and A gives engines delimited passage boundaries they can lift wholesale, which is precisely the format synthesis layers prefer.

Weeks 3 to 4: fix the entity

The second phase repaired brand facts. We rewrote the About page to state plainly what the company is, what category it operates in, when it was founded, by whom, and where it is based. Then we reconciled every external profile — the company's own social bios, its review-site listings, its database entries — to one canonical description. Answer engines rely on these facts to separate a brand from similarly named companies, and inconsistency forces a guess that often goes wrong.

We also confirmed the site was crawlable by the right agents. The robots file was checked for GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, and Google-Extended, and none were blocked. A single overzealous disallow line here can quietly zero out an entire program.

Weeks 5 to 6: build consensus off-site

With the site now extractable and the entity clean, we turned to consensus. The team pursued genuine third-party statements on the corpora that actually fed the category's AI answers — accurate customer reviews on the platforms engines cite, honest participation in the community threads where the category is discussed, and inclusion in independent comparison content. No astroturf: fabricated mentions get filtered and erode the trust the whole strategy depends on.

Weeks 7 to 8: re-scan and iterate

At day 60 we re-ran the original twenty prompts under identical conditions and compared against the Week 0 baseline. The subject had moved from absent to named in a clear majority of answers on at least two engines, with the restructured pages appearing as cited sources. The gaps that remained pointed directly at next month's work.

| Phase | Days | Primary action | Metric to watch | | --- | --- | --- | --- | | Baseline | 0 | Record citation share across 20 prompts | Starting share (near zero) | | Extractability | 1 to 14 | Rewrite top 10 pages, add schema | Passages with standalone answers | | Entity repair | 15 to 28 | Fix About page, reconcile profiles | Fact consistency across sources | | Off-site consensus | 29 to 42 | Earn reviews, forum presence, roundups | Trusted-source coverage | | Re-scan | 43 to 60 | Repeat prompt set, compare | Citation-share delta vs baseline |

Why 60 days and not 30

Two lags set the floor. Crawl-and-recrawl cycles mean structural changes take one to three weeks to show up in what engines retrieve, and consensus mentions need time to accumulate across independent sources. Pushing for results in 30 days usually means measuring before the pipeline has refreshed, which produces noisy, discouraging numbers. Sixty days lets the crawl catch up and gives the mention layer room to move.

What made it work, and what to copy

Three decisions did most of the lifting. First, sequencing: extractability before off-site. Earning mentions to pages that had no citable passage would have wasted the mentions. Second, measurement discipline: a fixed prompt set scanned under identical conditions, so every claim of progress was checked against a real control rather than a vibe. Third, honesty in the off-site work, because filtered fake mentions cost more than they earn.

The replicable core is a loop, not a launch. Restructure, fix the entity, build consensus, re-scan, repeat. If you want the instrumentation side handled from day zero, pair this with how to track AI mentions of your brand and run the baseline before you change a single page — the baseline is the part teams skip and the part that makes the case study real. You can see the tracking mechanics in Menra's citation tracking.

— The Menra Team