A first-party GEO case study: what 199,000 AI citations reveal about the playbook that works

Most writing about Generative Engine Optimization argues from theory. This one argues from a spreadsheet. It is a case study of a single publisher, guptadeepak.com, using its own Bing AI search citation data: 2,651 grounding queries and roughly 199,442 citations from one export in early July 2026. The short version: in the categories where the site publishes deep, dated, comparison-first content, it earns 40 to 50 percent of AI citations, and where it does not, it earns almost none. The gap is the lesson.

This is first-party data about my own properties, so read it as a worked example rather than an independent study. The value is not the specific numbers, it is the pattern they expose, which is reproducible on any site.

The data, and how to read it

The source is Bing's AI search "grounding query" report for guptadeepak.com. A grounding query is a query that triggered an AI answer, and citation share is the percentage of citations for that query that pointed to this domain. So a 50 percent share on a query means that, across the sources the AI assembled to answer it, half pointed here.

Three honest caveats before any number:

• One engine. This is Bing and Copilot's AI answer surface, not ChatGPT, Perplexity, Gemini, or Google AI Mode. Citation patterns diverge sharply across engines, so treat these shares as one engine's view, not the market's. See the GEO tooling market guide for why cross-engine coverage matters.
• One snapshot. A single export is a photograph, not a film. Citation share drifts month to month, so a snapshot cannot show trend.
• Correlation, not a controlled experiment. The site wins these queries and it publishes this kind of content, but nothing here isolates cause. It is consistent with the mechanisms in entity authority for AI engines, not proof of them.

With that framing, the numbers are still striking.

Where the site wins, and by how much

Grouping the queries by topic and weighting each query's share by its citation volume gives a clean picture of category-level performance.

At the individual-query level the wins are sharper still. Representative examples, each drawn from the export:

None of these are branded queries for the site. They are generic buyer questions, and the AI is choosing this domain as a top source roughly half the time.

The pattern behind the wins

Six characteristics separate the winning categories from the losing ones. Together they are the playbook.

1. Comparison-first structure

Every high-share query is a shortlist question ("top X", "best X", "X vs Y"). The pages that win them are structured as comparisons: a ranked list, a quick-comparison table, per-vendor pros and cons, and an explicit verdict. AI answers to shortlist questions are themselves shortlists, so a page already in that shape is trivially easy to extract and cite. Prose essays on the same topic do not win these queries.

2. Category depth, not a single page

The site does not have one IAM page, it has a cluster: IAM, CIAM, PAM, IGA, identity lifecycle, cloud directory, ITDR, and a whole CIAM knowledge portal beside them. Google's query fan-out breaks one question into many sub-queries, and Bing does something similar. A category covered by a dozen interlinked pages answers more of those sub-queries than a category covered by one, which is exactly why the deep categories win and the thin ones do not.

3. Direct, extractable answers with honest limitations

The winning pages lead with the answer and include the parts most marketing content omits: honest weaknesses, "not for you if" notes, and neutral verdicts. AI engines appear to prefer sources that read as assessment rather than advertisement. This matches the citation-worthy content patterns and the E-E-A-T signals covered elsewhere on this portal.

4. Freshness, made explicit

The strong pages carry a year in the title and a visible last-updated date, and they get re-dated when reviewed. Queries increasingly carry a year ("...2026"), and a page whose title and metadata match reads as current. Staleness is the main way a category-leading page loses ground, which is why keeping the year current is a maintenance task, not a one-time act.

5. Vendor-neutral framing

The site sells nothing on these pages. It names the real leaders, including competitors to the author's own products, and discloses conflicts where they exist. Neutrality is not only an ethics choice, it is a citation strategy: an engine synthesizing a balanced answer reaches for balanced sources.

6. Structured data and entity clarity underneath

Each comparison ships schema.org markup (ItemList, FAQ, Organization), and the brand's entity is anchored consistently across the network. This is infrastructure rather than a lever, and its exact contribution is unprovable from this data, but it is present on every winning page. The honest version of that argument is in homepage entity strategy for AEO.

Where the site loses, and why it is the same lesson

The losing categories confirm the mechanism by inverting it.

• AEO and GEO meta-topics: about 12 percent. For a portal that exists to teach GEO, being under-cited on GEO is the sharpest possible finding. The cause was simply thin coverage of those specific queries relative to the security categories. The fix was to publish the depth, which is the work this very guide is part of.
• Off-topic hardware terms ("gpu", "npu", "cpu"): near zero. The site has no business ranking for graphics-card queries and does not. Correctly so.

Both cases point the same way: citation share tracks published depth on a topic, not the size or authority of the brand in the abstract. A well-known author gets cited where they have done the work and ignored where they have not.

The playbook you can copy

Reduced to steps, the pattern is a program any B2B site can run.

• Pick a small number of categories you can genuinely own. Depth beats breadth; three deep categories outperform thirty shallow ones.
• Build the comparison layer first. For each category ship a "top N" ranked page and the head-to-head "X vs Y" pages buyers actually search, each with a table, per-option verdicts, and honest weaknesses.
• Add the definition and entity layer. Support the comparisons with glossary and explainer pages so you answer the fan-out sub-questions, not just the head term.
• Date everything and keep it current. Put the year in the title, show a last-updated date, and re-review on a schedule.
• Stay neutral and cite your sources. Name the real leaders, disclose conflicts, and write assessments rather than ads.
• Measure citation share and iterate. Instrument the categories, watch share by query and engine, and treat gaps as the next content backlog. The method is the AEO and GEO experimentation roadmap, and the metrics are in measuring AI visibility.

What would make this stronger

I hold this evidence loosely on purpose. A single engine and a single snapshot of my own data cannot carry a general claim. What would strengthen it: the same analysis across ChatGPT, Perplexity, Gemini, and Google AI Mode; a longitudinal series instead of one export; and controlled before-and-after tests on individual pages to move from correlation toward cause. Measuring citation share across engines is exactly what AI-visibility platforms do, including GrackerAI, the author's own company, disclosed here per the methodology.

The honest headline is narrow and still useful: on one major AI engine, a vendor-neutral B2B network earns close to half of all citations in the categories where it publishes deep, dated, comparison-first content, and almost none where it does not. If that pattern holds on your categories the way it holds on these, the playbook above is the whole game.