Independent research · AI source discovery
The Gay Creative Case Study: From AI Source Discovery to a Publisher Shortlist
What one independent creative practice revealed about the distance between finding relevant sources and identifying sensible editorial opportunities.
- Research collection
- 7 September 2026 UTC
- Matched research themes
- 10
- Source observations
- 957
- Shortlisted page-target records
- 40
On 7 September 2026, Kojable studied The Gay Creative, a Brooklyn creative practice, using Claude Sonnet 5 and Gemini 3.7 Flash across ten matched research themes. The runs recorded 957 source observations (733 Claude, 224 Gemini). Their normalised preferred-URL sets shared 24 of 766 distinct identities—3.13% exact URL overlap. The downstream pipeline separately shortlisted 40 page-target records (12 Claude, 28 Gemini); a later workbook proposed 14 editorial examples and ten content ideas. These were discovery, qualification and planning outputs—not measured AI citations, accepted placements or business outcomes.
The interesting part of the case is not simply that the two systems found different material. It is that the larger list did not become the larger shortlist, and that the shortlist still needed a different kind of editorial judgement before anyone could sensibly act on it.
Case question: How did AI-assisted source discovery become a reviewable publisher and creator shortlist for The Gay Creative?
At a glance: what the study counted
| Recorded research object | Claude | Gemini |
|---|---|---|
| Completed research tasks | 10 | 10 |
| Step 1 source observations | 733 | 224 |
| Source observations with a saved preferred URL | 731 | 203 |
| Distinct normalised preferred-URL identities | 625 | 165 |
| Step 2 sources classified candidate created | 19 | 30 |
| Shortlisted page-target records | 12 | 28 |
How to read this table: A source observation is one entry in a research task's saved source array; a source may appear more than once. A normalised preferred-URL identity is a deduplicated URL under the study's declared identity rules, not necessarily a unique publisher or article. A candidate-created disposition is a source-level classification. A shortlisted page target is a separate saved pipeline record. The 14 examples in the later editorial workbook are different again. These rows do not form a verified conversion funnel.
Evidence: Kojable's audit of the archived 7 September 2026 run manifests, source inventories, Step 2 classification outputs and publication shortlists. The underlying archive was checked by the research owner; the complete raw archive was not reprocessed in preparing this article.
What was Kojable researching for The Gay Creative?
The Gay Creative is the creative practice of Ezequiel Consoli, whose official About page describes an art-direction and creative-leadership practice based in Williamsburg, Brooklyn. Kojable carried out this research independently. The case is about the information landscape around that practice—not a client testimonial or a claim that a publisher or model endorsed it.
The starting point was a buyer problem. A strong portfolio can show the quality and character of creative work without settling every question a brand or agency faces when deciding how to hire an independent creative lead.
Consider two prospective buyers. One needs a senior person to shape a campaign idea while an existing team handles production. Another needs a larger delivery organisation, ongoing resources and a more extensive set of services. Both might search for “creative direction”, but they are trying to make different decisions. A source that helps with one question may offer little for the other.
The study therefore worked from ten predefined research themes rather than one broad “who is the best creative director?” query:
| Buyer-question group | Themes covered in the ten-task research set |
|---|---|
| Finding and evaluating independent expertise | New York freelance creative/art-direction options; choosing between an independent lead, agency, internal hire or marketplace; verifying experience and fit |
| Defining an engagement | Scope and budget; onboarding a freelance creative lead into an existing team |
| Understanding creative responsibilities | Leadership for integrated campaigns and launches; carrying visual identity into campaign execution |
| Reading professional evidence | Creative portfolios and the presentation of project roles |
| Understanding cultural and social work | Inclusive campaign leadership; social-first creative work and cultural relevance |
Those themes were research inputs. They are not ten statistically measured gaps in the creative-services market, nor do they establish how frequently buyers ask each question.
How did the September 2026 discovery and qualification process work?
The archived runs were completed on 7 September 2026 UTC. Each provider covered the same ten task IDs, P01–P10, once. They did not run a repeated, controlled model benchmark.
The process had four distinct parts:
- Discover sources. Claude Sonnet 5 used web search; Gemini 3.7 Flash used Google Search grounding. Each saved research task returned source entries for the specified question. Their search exposure and returned formats differed.
- Resolve URL identities. The audit used each source's saved preferred URL, set aside blank values for the URL comparison, applied conservative normalisation and deduplicated within each provider's ten-task set.
- Classify and shortlist. Both providers' research outputs then went through a common Gemini 3.7 Flash-based Step 2 classifier, with structured output and medium thinking. This second stage did not conduct a new source search. It assigned source dispositions and produced separate page-target shortlist records.
- Make an editorial assessment. A subsequent workbook documented candidate pages, why they might matter, proposed educational angles, reasons for caution and ten potential content ideas.
The distinction is crucial: the two AI systems were used for independent source research. Kojable was not measuring a panel of ordinary buyer-facing AI answers and the citations displayed in those answers. An entry in this research inventory is not proof that ChatGPT, Claude, Gemini or Google AI Mode recommended The Gay Creative, cited that URL to a buyer or incorporated the page into a final answer.
That separation is also described in Kojable's AI source research measurement framework. The framework was published after these September runs; this case is a retrospective illustration of its independent-research track, not evidence that applying the later framework caused an improvement.
What did the Claude and Gemini runs actually record?
The source inventories were markedly different. Claude recorded 733 observations across its ten tasks. Gemini recorded 224. Together, those are 957 source observations, not 957 unique pages, independent publishers or AI citations.
The URL data narrowed those counts in a different way. Two Claude observations and 21 Gemini observations had no saved preferred URL, leaving 731 and 203 entries with one. After conservative URL normalisation and within-provider deduplication, the sets contained 625 Claude identities and 165 Gemini identities.
The later classification stage identified 19 Claude and 30 Gemini source observations with a candidate-created disposition. The saved publication shortlists, however, contained 12 Claude and 28 Gemini page-target records, 40 in total. Those numbers describe different objects and should not be divided into one another as though every source observation followed a measured one-to-one path to an accepted opportunity.
For a research owner, the practical result is that each stage answers a different question:
- Discovery: What sources did the configured research system return?
- Identity: Which saved preferred URLs are distinct under the chosen rule?
- Disposition: How did the common policy classify each source observation?
- Shortlist: Which page-target records did the application retain for review?
- Editorial assessment: What, if anything, might be useful to investigate next?
The study preserved those questions separately. That is what made the result interpretable.
Why did the larger discovery inventory produce fewer shortlisted pages?
This was the clearest contrast in the saved results: Claude returned more than three times as many source observations, while Gemini's run produced more than twice as many page-target shortlist records under the same downstream classification process.
It would be tempting to call that a win for one model. The study does not support that conclusion.
The discovery systems were exposed to different retrieval processes. The source arrays were not equivalent-sized random samples from a shared web corpus. A large list can contain background references, owned pages, competitors, duplicates or items for which the evidence needed to assess an editorial opportunity is missing. A smaller list can happen to include more pages that fit a specific downstream policy. The common Step 2 classifier then applies the same application rules to the material it receives; it cannot make the initial retrieval conditions identical.
The observation is narrower and more useful: in this particular pair of runs, more discovered material did not correspond to more retained page-target records. Counting source entries alone would have hidden that reversal.
How much did the two discovered URL sets overlap?
Kojable's archive audit identified 625 distinct normalised preferred URLs for Claude and 165 for Gemini. Exactly 24 URL identities appeared in both sets. Their combined union therefore contained 766 identities:
625 + 165 − 24 = 766
The Jaccard overlap for the two URL sets was:
24 shared identities ÷ 766 identities in the union = 3.13%
The unit here is an exact normalised preferred-URL identity across the pooled ten tasks, not a topic, organisation, creator, citation event or model recommendation. The comparison excludes source observations with blank preferred URLs. The normalisation removes fragments, a leading “www.” and recognised tracking parameters while retaining meaningful query parameters, such as the video identifier in a YouTube URL. It does not merge publisher aliases, alternative video URL forms, redirects or different articles on the same subject.
That means the measured overlap is low under this specific identity rule. It does not mean the two systems had only 3.13% thematic agreement, that their publisher lists barely intersected, or that their final AI-answer citations would overlap by the same amount. The saved archive did not separately verify an exact-overlap figure for the shortlisted 12 and 28 page targets, so no such figure is claimed here.
The case suggests a research question worth asking when source pools differ: does the second pool add genuinely useful perspectives or eligible editorial contexts? The URL intersection alone cannot answer it.
How did source eligibility and review change the shortlist?
Topical relevance and eligibility for an independent editorial contribution are not the same judgement.
A company's own portfolio can be indispensable evidence about its work while being unsuitable as an independent publisher target. A competitor's service page may help define an alternative without being an appropriate contributor to contact. A relevant video might still need a confirmed owner, accurate date and clear rights before anyone proposes using it.
The Step 2 source dispositions illustrate how the saved process distinguished these situations:
| Selected source-observation disposition | Claude | Gemini |
|---|---|---|
| Reference only | 387 | 85 |
| Business-owned or competitor-owned | 100 | 23 |
| Insufficient page evidence | 75 | 34 |
| Manual-review candidate | 35 | 15 |
| Candidate created | 19 | 30 |
| Subtotal: five selected categories | 616 | 187 |
| Other saved dispositions | 117 | 37 |
| All source observations | 733 | 224 |
Counting note: These are source-observation dispositions, not distinct URLs, publishers or page-target records. The ownership row combines 42 business-owned + 58 competitor-owned observations for Claude and 2 + 21 for Gemini. The selected five-category subtotals are not exhaustive; the remaining 117 Claude and 37 Gemini observations were assigned other saved labels, including not topically relevant, non-addressable, invalid, duplicate or commissioned-gap basis. Every observed source appears in the full 733/224 disposition denominators.
The 35/15 manual-review counts belong at this source-disposition stage. They are not “page-target records requiring manual review”. Likewise, 19/30 candidate-created dispositions are not interchangeable with the 12/28 shortlisted page-target records.
The point of this separation was not to dismiss every source that failed an outreach test. A reference-only page may still be valuable for understanding what a buyer can learn. It simply supports a different decision from a page that a researcher might put forward for human editorial review.
What did the editorial review propose?
The next output was a dated editorial opportunity workbook, prepared on 8 September 2026. It contained 14 page-level examples under four review statuses, with each example accompanied by a proposed angle and a verification guardrail.
| Editorial workbook status | Examples | Meaning in this case |
|---|---|---|
| Best opportunity | 4 | Apparent strong topic fit and a plausible reason to investigate a contribution |
| Good opportunity | 6 | Relevant, but information such as rates, claims or scope needs more care |
| Follow Up | 1 | Existing topical coverage merited a further check or possible conversation |
| Check First | 3 | Hold any approach until an explicit verification issue is resolved |
| Total | 14 | Curated editorial examples, not confirmed placements |
Two entries show why the qualitative review mattered.
The workbook classified Jack Appleby's article about relying less on social-media trend content as a Best example. The possible fit was educational: adding a practical discussion about starting with human insight before adapting ideas to a platform. It did not mean that the author wanted a contribution, endorsed The Gay Creative or had been approached.
It also classified GAYLETTER's existing coverage of #PRIDETRAIN as Follow Up. The published story provided a relevant context involving the creative practice's wider project history, but the workbook still required exact collaborator roles to be checked. The distinction matters: existing coverage can justify looking more closely without licensing an inflated claim about who did what—or proving that a new publication opportunity exists.
The Check First group took the same care a step further. Its notes included confirming a page owner's contact route, checking whether an article still matched the proposed topic and verifying a video channel's ownership before contact. A credible research list should be able to say “not ready”.
The fourteen examples are not fourteen unique creators: one named creator appears against two different pages. They are also not a mathematically verified selection rate from the forty machine-shortlisted records. The retained evidence does not provide a complete record-by-record join between those two presentation stages.
Ten content ideas, grouped around five buyer decisions
The workbook separately proposed ten potential content assets. They were not completed publications. For the case study, their themes can be synthesised into five families of buyer questions:
| Buyer information need | Example of a proposed response |
|---|---|
| Choose the right creative arrangement | A decision guide comparing an independent senior lead, agency, internal hire and marketplace, plus a New York-specific hiring guide |
| Set scope and handoffs clearly | A campaign-scope and budget worksheet, an onboarding playbook for a freelance Associate Creative Director, and a responsibility map for integrated campaigns |
| Verify experience and professional roles | A freelance creative-partner hiring checklist and a portfolio-redesign brief that preserves truthful role attribution |
| Carry creative systems into execution | A guide to taking visual identity into real campaign and channel work |
| Develop culturally informed creative work | An authority-and-approval framework for inclusive campaigns and a guide to social-first ideas without reflexive trend chasing |
This five-family grouping is an editorial synthesis of the ten workbook proposals, not a separate measured content-gap taxonomy. The research raised plausible information needs around role clarity, scope, evidence, cultural context and collaboration. It did not prove that these needs were absent across the entire market or that publishing the ideas would change AI answers.
What should an expertise-led service company take from this case?
For Kojable, the central lesson is about the quality of the decision after source discovery.
The Gay Creative's situation is specific, but the distinction it reveals is relevant to other expertise-led businesses. A consultancy may have strong project credentials without clearly explaining who led the work. A specialist agency may publish detailed services without helping a buyer decide when it is a good fit. An independent professional may have impressive third-party coverage that does not answer the practical questions buyers need resolved.
The case suggests three different responses to three different findings.
If the missing information concerns the practice itself, such as services, scope, decision rights or a person's exact role, the appropriate first response may be clearer owned evidence. An independent article should not be asked to stand in for factual gaps that the business can resolve directly.
If a credible external page addresses the buyer's question but leaves a useful angle unexplored, it may be worth researching as an independent editorial context. That requires checking audience, author identity, freshness, evidence, publication fit and the owner's independent judgement. Mere topical similarity is too weak.
If an apparent opportunity relies on uncertain facts, such as public rates, individual campaign credits, awards, reach or contact ownership, the useful output may simply be a verification task. That is not a failure of the process. It prevents an attractive-looking list from turning into poorly grounded outreach.
None of these responses was tested here as an intervention. They are decisions informed by one documented case, not a causal finding about how to gain rankings, citations or revenue.
What the study cannot establish — and how its measurements can be understood
The bounds of this case are part of its result.
- One practice, one period. The study concerns The Gay Creative and ten chosen research themes. It contains one completed Claude discovery run and one completed Gemini discovery run from 7 September 2026, with no repetitions. It is descriptive and exploratory.
- Different discovery systems, shared downstream classification. Claude web search and Gemini search grounding were not identical retrieval conditions. Both source sets were assessed by the common Gemini-backed Step 2 process. The results do not rank the models' general quality.
- Exact URL identity is narrower than shared subject matter. The 24/766 measure uses saved preferred URLs after conservative normalisation. It does not merge related pages, URLs that redirect to the same destination, creators or publisher organisations.
- The saved files are not a perfect replay bundle. Kojable's archive audit reported manifest-hash matches for 40 raw and normalised Step 1 files, two policy snapshots and two Step 2 response files. The original aggregate files were missing, and both runs recorded uncommitted source changes. Those integrity checks support the historical record; they do not guarantee full code-state reconstruction or identical future model output.
- An independent source inventory is not an AI-answer citation panel. This study did not measure whether a final answer cited these pages, whether a source shaped an AI recommendation, or whether any ranking, indexing or attribution stage changed.
- Proposals are not outcomes. The workbook recorded potential educational approaches and content ideas. It did not measure publisher contact, acceptance, published contributions, subsequent AI-answer effects, qualified leads or revenue.
For readers comparing the numerical findings, the basic measurement contract is:
- Observation denominator: all saved Step 1 source-array entries across P01–P10 per provider (733 and 224), with repeats retained.
- URL identity eligibility: only entries with saved preferred URLs (731 and 203), normalised and deduplicated within provider (625 and 165 distinct identities).
- Overlap numerator and denominator: 24 exact shared identities out of 766 in the cross-provider union, giving 3.13%.
- Classification denominator: all 733/224 observations, including the statuses outside the five selected categories.
- Shortlist unit: saved page-target records (12 and 28), distinct from the 19/30 candidate-created source dispositions.
- Editorial unit: 14 dated workbook page examples and ten separate proposed content ideas. No observed conversion ratio connects those counts to the other stages.
Kojable's research-owner audit established the reported historical counts from the retained archive. The underlying raw records were not reanalysed independently for this publication draft. The detailed internal claim–source ledger records the source locations, calculation rules and exclusions.
Frequently asked questions
Did Claude find more sources than Gemini in this study?
Yes, as source observations in these specific archived runs. Claude recorded 733 entries and Gemini 224 across the same ten research themes. A source observation is an entry in a saved research array, not a unique page or an observed AI-answer citation. One paired study cannot establish general discovery superiority.
Why did Gemini have more shortlisted page targets despite fewer source observations?
The saved pipeline retained 28 Gemini page-target records and 12 Claude records, even though Claude's initial source array was larger. Different retrieval exposure and source composition can affect what the common eligibility process receives. The records demonstrate a stage-level reversal; they do not isolate a causal reason or prove that Gemini was generally better.
Were the discovered pages actually cited in AI answers?
That was not measured. The source arrays came from AI-assisted research with web search/grounding. The study did not capture and analyse a separate panel of ordinary buyer-facing final answers with citation supports. It would be incorrect to call the 957 observations “957 AI citations” or to treat absence of such a panel as zero citations.
What does the 3.13% URL overlap measure?
It is 24 exact shared normalised preferred-URL identities divided by 766 identities in the union of the two ten-task discovery sets. Blank preferred URLs were excluded, and each provider's URLs were deduplicated. The result does not measure publisher, author, topic or final-answer citation overlap.
How do shortlisted page targets differ from the fourteen editorial examples?
The 40 page-target records were saved by the machine-assisted shortlist stage: 12 Claude and 28 Gemini. The 14 workbook examples were later presented with human-facing Best, Good, Follow Up and Check First labels and verification instructions. They have different units and no fully documented record-by-record mapping that would justify a conversion rate.
Were any publishers contacted or content ideas published?
Those outcomes were not measured by this study. The workbook documents proposed approaches and ten potential content ideas. Its labels do not prove that someone was contacted, consented, contributed, published or produced an AI-answer or commercial effect. The research archive examined here is a discovery-and-planning record.
Can the same research approach apply to other professional-services businesses?
The measurement distinctions can guide another study: define buyer questions, record source observations, resolve identities, assess evidence and eligibility separately, then distinguish proposals from completed actions. The specific publisher candidates and content ideas would need to be researched afresh for the business, sector and buyer decision. This case does not validate the method's effectiveness across sectors.
Conclusion: the shortlist is not the end of the research
The Gay Creative case began with 957 recorded source observations. It exposed very different URL pools, a reversal between discovery volume and page-target shortlisting, and a later editorial layer that included not only promising examples but explicit reasons to pause.
The key decision was not “which AI returned more?” It was which research objects had actually been measured, which sources were appropriate for the question, and which proposed actions still needed evidence.
For a company trying to be accurately understood by buyers—including during AI-mediated discovery—that discipline matters. Better decisions begin with the right questions and a clear account of what the evidence can and cannot say.
This case illustrates the independent-research side of Kojable's AI source research measurement framework. Kojable's Creator & Publisher Intelligence capability addresses the related task of investigating relevant independent voices and preparing evidence-led next steps for human review. Neither research discovery nor a proposed contribution guarantees publication, indexing, citation or a recommendation.
Research provenance and source notes
First-party research: Independently conducted by Kojable for The Gay Creative. Two saved ten-task discovery runs dated 7 September 2026 UTC, using Claude Sonnet 5 (archived model ID: claude-sonnet-5) with web search and Gemini 3.7 Flash (archived model ID: gemini-3.7-flash) with search grounding. Common downstream Gemini 3.7 Flash structured classification, source dispositions and shortlist exports. Metrics were established by Kojable's archive audit of the saved run records; that audit also documented available-file hash integrity and its reproducibility limitations.
Editorial source: The 8 September 2026 The Gay Creative — AI visibility, outreach opportunities and content ideas workbook, directly checked for fourteen page-level examples, four review statuses and ten proposed content ideas. The workbook is a planning document, not evidence of completed contact or placement.
Public contextual sources: The Gay Creative — About; Kojable's AI source research measurement framework; Kojable's Creator & Publisher Intelligence. The two illustrative outside-page examples are linked beside the relevant discussion above. Public third-party pages are used as illustrations of the workbook's editorial judgement, not as validation of the archived quantitative metrics.