Original Research

Content Gaps in B2B Technology: Patterns from Nine Research Audits

Published By Piush Vaish

Across nine anonymised B2B technology research audits, Kojable reviewed 90 structured buyer-question records and documented 172 source-target records and 105 proposed content-gap records. A qualitative synthesis identified seven recurring evidence themes, from product boundaries and technical validation to deployment, economics and security. These findings describe what the audits recommended investigating or improving. They do not establish 105 verified missing pages, 172 confirmed AI citations, or evidence that any recommendation changed an AI answer or generated revenue.

Key findings

  • The nine audits contained 90 buyer-question records, 172 source-target records and 105 proposed content-gap records. Each number describes a different unit of analysis.

  • The recommendations could be organised into seven recurring evidence themes: category and product boundaries; comparative evaluation; reproducible technical proof; deployment and migration; cost and economics; security and governance; and operational workflows and outcomes.

  • The underlying work was organised around ten structured questions per audit. Recurrence may therefore reflect the audit design as well as the concerns examined in each technology category. It is not a market-wide frequency ranking.

  • The workbooks contain useful AI-answer summaries and source research, but their model and citation fields are not standardised enough to establish why an AI system selected or omitted a source.

The most useful question arising from this work is not simply, “What should we publish next?” It is: What evidence would make a particular technical buying question answerable, comparable and verifiable?

  • B2B evidence requirements
  • Content-gap research
  • 18 min read

1. What counts as a content gap in this research?

In this study, a proposed content gap is a recorded recommendation to develop, clarify, validate or better connect information relevant to a defined buyer question. It might call for a comparison framework, a technical test, a migration procedure or more transparent pricing assumptions.

That is narrower than saying the information is absent from the internet. It is also different from saying an AI system failed to retrieve a document.

A recommendation can arise because a question is difficult to answer with the available evidence, because existing information appears fragmented, or because a researcher believes a more useful form of documentation is warranted. Confirming that a vendor actually lacks the information would require a separate, documented review of its current content and supporting sources.

This distinction matters when teams investigate AI answers. An omission in an answer is something that can be observed. The reason for that omission may remain unknown. Kojable's separate guide to diagnosing missing proof in AI answers addresses that more demanding verification question.

2. Research scope and method

We conducted a retrospective review of nine historical B2B technology audit workbooks, with dated records spanning July to September 2026. The audits covered different technical markets and buying situations. We have withheld client and product names, identifying source URLs and campaign-specific details from the published examples.

Each workbook contained four comparable analytical components:

  1. Campaign summary: the scope and high-level research interpretation.

  2. Tracked questions: ten structured buyer questions and associated AI-answer summaries.

  3. Citation targets: source entries researched and shortlisted in relation to one or more questions.

  4. Content gaps: proposed subjects, coverage requirements and validation needs.

Some workbooks contained additional review and verification fields. Their definitions were not fully consistent across audits. Certain dates denoted a reviewed or rebuilt workbook rather than a newly executed model run, so the archive should not be treated as a single synchronised, controlled experiment.

For this retrospective, we counted the recorded rows in the three relevant sheets and used their question references to inspect relationships between questions, targets and proposed gaps. We then grouped the question themes and proposed coverage requirements into seven qualitative categories. This is an interpretive synthesis of existing research records, not a blinded coding exercise, a buyer survey or a statistical estimate of market prevalence.

No new AI-answer runs were conducted for this article. The two separate consumer-facing studies available in the wider research archive were excluded.

For Kojable's broader approach to distinguishing discovered sources, observed citations and proposed actions, see the AI Source Research Measurement Framework.

3. Measurement dictionary: three records, three meanings

Measurement dictionary for the three record types
Record typeRecorded countUnit and interpretationWhat it does not prove
Buyer-question record90One tracked question within an audit; ten per auditNinety unique buyers, independent searches or representative market demand
Source-target record172One source entry selected for research or further evaluation, sometimes associated with multiple questionsA unique verified citation, publisher commitment, placement or prospect
Proposed content-gap record105One proposed content or evidence improvement recorded by the auditA verified missing page, completed intervention or demonstrated business result

The row counts are reported without claiming cross-workbook URL deduplication or independent verification of the suggested gaps. They should not be summed into one measure of opportunities.

The question identifiers also allow a limited relationship check. 77 of the 90 question IDs appear in at least one source-target reference, while 89 of the 90 appear in at least one content-gap reference. These figures describe recorded links in the workbooks, not the proportion of questions successfully answered, cited or commercially addressed. A single question can connect to several targets or recommendations.

This measurement discipline is important because the same word—source—can refer to very different things: a URL researchers discover, a publication they shortlist, a page an AI answer cites, or a page whose content has been independently checked against an answer claim. Each requires different evidence.

4. Seven recurring B2B technology evidence themes

The seven categories below organise the needs recorded in the audits. They are not ordered by prevalence or importance. Because the ten-question structure was partly shared across audits, the categories should be read as recurring themes within this research process, not as the seven statistically most common gaps among all B2B companies.

Seven evidence themes and representative buying questions
Evidence themeRepresentative buying questionEvidence an audit might propose
Category and product boundariesWhat does this product do, and where does its responsibility end?Scope definitions, architecture, supported use cases and limitations
Comparative evaluationHow should competing approaches be compared fairly?Matched criteria, consistent assumptions and trade-off documentation
Reproducible technical proofUnder what conditions does a technical claim hold?Test methods, benchmarks, validation protocols and failure cases
Deployment and migrationHow does a team implement, migrate or recover safely?Prerequisites, compatibility, rollout, rollback and operational runbooks
Cost and economicsWhat drives the actual cost and expected value?Total-cost models, inputs, commitments and sensitivity assumptions
Security and governanceWhich controls and limits govern the system?Threat boundaries, permissions, assurance records and applicable requirements
Operational workflows and outcomesWhat changes for the people running the system?End-to-end workflows, integrations, responsibilities and measurable outcomes

Category and product boundaries

Technical categories often overlap. Buyers need to understand whether a solution provides a complete operating layer, a specialised component, a service wrapper or something that must be combined with other tools. A useful answer needs to identify what is included, what remains the buyer's responsibility and where dependencies sit.

The audits proposed clearer category boundaries and capability explanations in several settings. A broad claim such as “end-to-end platform” is difficult to evaluate without a description of endpoints, responsibilities, supported environments and exclusions. The research does not establish that any named vendor lacked these details; it documents the kinds of clarity the audit process sought.

Comparative evaluation

Comparison is more demanding than placing product names in adjacent columns. Different approaches may have different service boundaries, operating assumptions, testing conditions and cost structures. A comparison can mislead even when each individual statement is technically true.

Audit recommendations therefore called for shared criteria, matched scenarios and explicit limitations. In technical markets, a credible comparison may need to distinguish a measured result from a product claim, and a supported feature from one available only through custom implementation or an adjacent service.

Reproducible technical proof

The question behind many technical claims is: Can another qualified person tell how the result was obtained?

Across the archive, proposed coverage included benchmark conditions, validation protocols, repeated tests, workload definitions, thresholds and failure handling. These are different forms of proof, but they share an important property: they make a claim inspectable rather than relying on an adjective such as fast, reliable or secure.

Reproducibility is not always possible in the same form. A software workload, an industrial measurement and a regulated system may require different controls. The common evidence need is to identify the test boundary, method and uncertainty.

Deployment, migration and recovery

Enterprise evaluation rarely ends when a feature looks attractive. The buyer must understand what it takes to deploy the system, integrate it into existing operations and recover when something goes wrong.

The audits repeatedly proposed implementation detail: environmental prerequisites, compatibility decisions, staged rollout, migration sequencing, observability, fallback procedures and rollback. These recommendations address the gap between a capability description and an operationally credible plan.

Cost and economics

Headline prices can be poor substitutes for total cost. The full comparison may depend on usage patterns, infrastructure, storage, support, engineering effort, utilisation, commitments and recovery overhead. For a technology buyer, those assumptions can change the decision.

The workbooks proposed cost models tied to defined workloads and operating conditions, rather than unsupported claims about savings. That is a request for decision-quality evidence. It is not proof that the proposed option was cheaper or that a published calculator would have changed a purchase.

Security and governance

Security questions become difficult when a general assurance statement is used to answer a specific architecture or operating question. Buyers may need to know who can access what, where sensitive information crosses a boundary, which controls are tested and which responsibilities remain with their own team.

The audit recommendations included control maps, threat assumptions, permission boundaries, verification methods and operational responses. A useful document should distinguish implemented controls, stated policies and claims still requiring validation. A list of security terms alone is not equivalent to evidence that those controls work in a buyer's setting.

Operational workflows and outcomes

Finally, buyers often need to see how a product works in a complete process. A feature demonstration may omit the hand-offs, data movement, exception handling, ownership and monitoring that determine whether the system is usable in practice.

The proposed content included workflow walkthroughs, integration detail, operating models and clearer outcome measures. The emphasis was on moving from a feature statement to a process a buyer can evaluate. The archive records requests for such evidence; it does not measure realised productivity, reliability or commercial gains.

5. Three anonymised examples from the audit records

The following examples are paraphrased from linked buyer-question and content-gap records in three different audits. They illustrate the recommendation process; they are not quotations from customers or verified examples of companies failing to publish adequate evidence.

Example A: Comparing the real cost of technical infrastructure

One audit tracked a buyer question about comparing the cost of alternative infrastructure approaches. The related content recommendation went beyond headline capacity or hourly rates. It called for matched workloads, dated commercial assumptions, storage and network charges, support, commitments, idle time and operating effort.

The underlying evidence need was straightforward: without comparable inputs, a cost comparison could be accurate in parts yet misleading as a decision aid.

What was observed: an audit question linked to a proposed workload-level total-cost analysis. What was not established: that one supplier was cheaper, that the analysis was later published, or that it changed an AI answer.

Example B: Turning a technical signal into an operational decision

A second audit examined how operators could identify changes in the condition of a technical system before those changes affected reliability. The proposed content linked condition thresholds, repeat-test logic, independent confirmation and maintenance actions.

The opportunity was not merely to explain what could be measured. It was to document how an operator should distinguish a meaningful signal from noise and decide what to do next, subject to validated thresholds and system-specific assumptions.

What was observed: a recommendation for a test-to-action workflow. What was not established: that the thresholds were validated by the retrospective research or that the proposed workflow had been tested in production.

Example C: Planning a safe software rollout and recovery

In a third audit, a buyer question concerned releasing software changes gradually and recovering if an update introduced crashes or incompatibility. The proposed evidence covered compatibility checks, staged cohorts, telemetry thresholds, pause conditions, rollback behaviour and fallback arrangements.

The content need was operational specificity. A statement that updates can be deployed quickly does not answer how an organisation limits the impact of a bad release.

What was observed: a question linked to a proposed release-and-recovery playbook. What was not established: that the vendor lacked such a playbook, or that the recommended safeguards were implemented and independently verified.

Together, these examples show why a content-gap record can be more useful when it describes the decision and its required evidence rather than simply naming a topic for an article.

6. A source target is not a verified AI citation

The nine workbooks contain 172 source-target records. Some identify potentially relevant publishers or technical pages; some include prioritisation, source-review or grounding annotations. But the fields are not uniform enough to turn the total into a nine-audit count of confirmed AI citations.

Only 59 of the 172 target records include a cites field. Of those, 39 are marked “Not measured” and 20 carry a cited flag within one audit. Other workbooks use different fields or lack an equivalent. These entries cannot be combined into a comparable citation rate without reviewing the underlying answer evidence and applying the same eligibility and verification rules across all nine cases.

A useful measurement dictionary separates at least four statuses:

  1. Discovered source: a page or publisher identified through research.

  2. Shortlisted target: a source selected for further review because its topic or audience appears relevant.

  3. Observed citation: a source actually referenced in a recorded AI answer under specified conditions.

  4. Verified support: an identified source that has been checked to support the particular claim attributed to it.

The distinctions are not administrative details. A publication may be relevant without having been cited. A citation may be observable without establishing that the source supports a material claim. Kojable's source-identity research examines why claims about exact pages need stronger identification than a source label alone.

This is also the distinction carried into Creator & Publisher Intelligence: sources observed in AI answers, independently researched publishers and reviewed contribution opportunities belong to separate stages of work. None is automatically a partnership, placement or future citation.

7. What the audit archive can tell us about AI systems

Every tracked question in this sample has an AI-answer summary. That makes the archive useful for understanding how the researchers interpreted the answers available to them. It does not make the summaries equivalent to a standardised set of raw model responses, complete citation supports or retrieval logs.

For this retrospective, the five-stage pipeline diagnosis is therefore limited:

What the audit records can establish at each AI-answer stage
AI-answer stageWhat can be established from these records
Discoverability and indexingThe recorded source URLs do not establish crawlability or indexing at the historical run time. Unknown.
RetrievalSome records carry grounding annotations, but the full candidate pools and retrieval conditions are not consistently preserved. Mostly unknown.
Selection and rerankingResearcher-assigned target ranks are not model selection ranks. Unknown.
SynthesisThe workbooks preserve answer summaries, not a uniformly comparable raw-response archive. Partially observable.
AttributionA subset has citation-status fields, but the definitions and supporting records are inconsistent. Partially observable.

These limits rule out several tempting conclusions. We cannot infer that one model systematically favoured a category of publisher, that an omitted company failed an indexing gate, or that a proposed content improvement would have changed retrieval or citation selection. The archive is not a controlled Claude-versus-Gemini comparison.

8. What B2B teams can take from these findings

The study offers a practical way to inspect a content plan without treating every proposed topic as an instruction to publish another page.

Start with a buyer question. Ask what answer would support a real evaluation, what evidence that answer requires and whether the current evidence has been verified. The appropriate next action might be to improve existing documentation, clarify a claim, add test conditions, create a comparison, repair a fragmented workflow explanation or investigate independent coverage. It could also be to make no change when adequate evidence already exists.

This is where the research connects to Kojable's broader See → Align → Prove → Compound approach. See captures what an AI system said and separates the observed gap from possible explanations. Align scopes a justified information change. Prove examines comparable subsequent answers and factual quality. Compound uses the findings to prioritise the next question or evidence gap. None of these steps guarantees a citation or recommendation.

The detailed planning decisions belong in B2B Content Strategy for AI Search, not in this research paper. Questions about identifying and assessing relevant independent publishers belong in Creator & Publisher Intelligence.

There is a wider reason to care about evidence quality. In its B2B Content Guide, Forrester described technology buyers' expectations for relevant and credible vendor information before and after purchase. That external buyer-preference research provides context; it does not validate the frequencies or outcomes in our nine-audit sample.

9. Limitations and reproducibility

These findings should be interpreted within the following boundaries.

Sample and selection. Nine historical research engagements are a convenience sample, not a random sample of B2B technology companies, buyers or AI answers. The results cannot establish market-wide prevalence.

Structured-question effects. Each audit contained ten questions following a broadly similar research structure. Some recurrence may be a consequence of what the audit process deliberately asked researchers to examine.

Qualitative synthesis. The seven themes are an organising interpretation of recorded questions and recommendations. They have not been presented as independently coded prevalence estimates, and no per-theme ranking is claimed.

Uneven model and source fields. Model names, grounding status, citation flags and verification details vary across the archive. Where records do not support a common denominator, no cross-model rate is calculated.

Recommendation versus confirmed absence. Content-gap records specify proposed coverage. This retrospective did not independently audit the full live content inventory for every company or certify that each suggested piece was missing.

Historical timing. Workbook dates fall between July and September 2026, with some identifying reviews or rebuilds. AI systems and public information can change. These records should not be treated as measurements of today's answers.

Privacy and auditability. The published synthesis removes client identities, identifying campaign data and source URLs. The private archive retains the underlying question and recommendation records for internal traceability, but full public row-level reproducibility is constrained by confidentiality.

No tested intervention or commercial outcome. This research did not measure publication, outreach acceptance, AI-answer improvement, leads, conversion or revenue attributable to the recommendations. It cannot establish causation.

The appropriate evidence claim is therefore limited but useful: these nine audits generated recurring categories of proposed content and proof requirements that can inform what a team chooses to investigate next.

Frequently asked questions

What is a B2B technology content gap?

In this research, it is a documented recommendation to improve the information or evidence associated with a technical buying question. It may concern category clarity, comparison, validation, deployment, cost, governance or operational use. A recommendation alone does not prove the relevant content is absent.

Does the study identify 105 missing pages?

No. It contains 105 proposed content-gap records. A record may describe a new page, an improvement to existing material or a broader evidence requirement. The number is not a count of verified missing URLs.

Were 172 sources cited by AI systems?

No such conclusion follows. The 172 figure counts source-target records in the audit workbooks. Citation status is only partially recorded and not standardised across the nine audits. A research shortlist is not an observed citation set.

Can the results tell us whether Claude or Gemini is more likely to cite a particular source?

No. Model metadata and citation definitions are not consistent enough for that comparison. A defensible model-overlap study needs matched questions, comparable observation windows, consistently resolved sources and clearly defined eligibility.

Can we generalise the seven themes to every B2B technology market?

They are useful categories for organising evidence needs, but their prevalence cannot be generalised from this sample. The structured audit design and choice of markets influence which themes appear.

Did the recommended content changes improve AI visibility or generate customers?

That was not measured in this retrospective. The workbooks contain research findings and proposed actions, not a verified before-and-after intervention study or revenue attribution dataset.

Conclusion

The nine audits reveal a consistent feature of the research process: useful content recommendations were framed around questions a technical buyer needed to answer, followed by the documentation, test conditions, comparisons or operational evidence that could make an answer more defensible.

That is a stronger starting point than counting topics or treating every uncited source as a publishing opportunity. But it is still a starting point. A proposed gap must be checked, an improvement must be justified, and any claimed effect on AI answers or business outcomes must be measured separately.

Kojable works on that distinction between what an answer shows, what the evidence supports, what might need to change, and what can be verified afterwards.

About the author

Piush VaishFounder and CEO of Kojable