Your competitive intelligence has a source problem

Start with two facts about the market you're analysing.

First: the claims are inflated, measurably. When Gartner examined the agentic AI vendor landscape, it estimated that of the thousands of companies marketing agentic capabilities, only around 130 were building the real thing. The rest were rebranding chatbots, assistants and RPA. The practice has a name now — agent washing. If you built a competitive feature matrix from vendor websites in that category, you would have produced a document describing marketing departments, not products.

Second: the secondary sources are degrading. A study published in The Lancet by Columbia researchers examined over 2.5 million papers and 97 million citations and found fabricated references rising sharply: 1 in 2,828 papers contained one or more in 2023, 1 in 458 by 2025, and a further multiple of that during the first weeks of 2026. If peer-reviewed literature is absorbing invented citations at that rate, the trade press, analyst summaries and market reports feeding your CI process — none of which have peer review — are not doing better.

It gets worse for one specific CI staple. Research on AI-generated fake reviews found that humans detect them at roughly chance level, and that LLMs did no better — both landed around 50%, no better than a coin flip. Review mining, one of the most trusted CI inputs, now rests on a corpus nobody can reliably authenticate.

So the problem in 2026 is not access to data. Data is abundant and cheap. The problem is that a growing share of it is unverifiable, and the tooling that promises to process more of it faster does nothing about that.

Which suggests a different organising principle for competitive intelligence in tech markets.

Rank signals by what they cost to fake

Here is the frame I'd use in place of a feature matrix. Every competitive signal has a cost to fake — how much it would cost a competitor to make it say something untrue. Weight your intelligence accordingly.

Free to fake (treat as marketing, not evidence):

  • Website feature lists and capability claims
  • "AI-powered" labelling of anything
  • Press releases and launch announcements
  • Self-reported performance statistics with no methodology
  • Testimonials without a named person and company
  • Anonymous reviews

Costly to fake (treat as evidence):

  • Job postings. The most under-used signal in tech CI. A company hiring three infrastructure engineers with streaming experience is telling you about its roadmap eighteen months early, and paying salaries to do it. Job specs also leak the stack, the team structure, and which problems are unsolved.
  • Pricing page changes. Price is where strategy becomes non-reversible. A new tier, a removed one, a shift from published to "contact us," a change in the metering unit — each reflects a decision someone defended internally.
  • Documentation and changelogs. What shipped, not what was announced. Docs also reveal limits: rate caps, supported auth methods, deprecated endpoints. Deprecations are especially informative — they show what didn't work.
  • Integration and partner lists. Where they've invested engineering time, and by omission, which ecosystems they've declined.
  • What they refuse to do. A vendor who won't self-host, won't touch regulated data, won't go below a contract size, or won't support a legacy system has drawn a boundary. Boundaries are architectural and expensive to move, which makes them the most durable competitive information available — and they mark where underserved customers live.
  • Hiring and departure patterns at senior level, funding, patent filings, and partnership moves. A biometric SDK launch means one thing; the same launch alongside three device-intelligence hires means something more specific.
  • Status pages and incident histories. Reliability is expensive to fake and rarely discussed in sales conversations.
  • Win-loss interviews with your own lost prospects. The single highest-value CI input most companies never systematise, because it's uncomfortable.

The principle generalises: watch what costs money, not what costs words. Anything a competitor spent budget, engineering time or contractual commitment on is a signal. Anything they typed is a hypothesis.

Why feature matrices mislead in tech markets specifically

Most CI output is a grid: competitors down one axis, features across the other, ticks in cells.

In mature categories that works, because claims are checkable and roughly standardised. In fast-moving tech categories it fails in three ways.

Claims aren't comparable. Two vendors both tick "self-healing tests" or "multi-agent orchestration" or "AI-powered analysis" while shipping fundamentally different things. The tick hides the difference that matters.

The matrix measures presence, not quality. A feature that exists and is bad scores identically to one that's excellent. Buyers experience the second axis; your matrix doesn't have it.

It rewards claiming. A competitor who overstates gets a fuller row. Your matrix converts their marketing aggression into apparent product superiority, and then someone builds a roadmap from it.

The replacement isn't a better matrix. It's a different question set:

  • What do their customers complain about specifically? Complaints are more reliable than feature lists, because nobody writes a fake complaint about accuracy on a workflow they don't use. (Verify where you can — see the fake-review problem above. Complaints in support forums and community channels are harder to manufacture than review-site posts.)
  • Who are they not built for? Most products serve the well-resourced case — clean data, technical buyer, existing team to operate it. The unserved market is usually the same problem in a messier context.
  • What are they measuring, and what aren't they? If every competitor reports the same visible metric, the gap is often in the metric nobody publishes. Support automation vendors publish deflection; few publish true resolution or repeat-contact rate. Outbound vendors publish meetings booked; few publish meeting-to-opportunity conversion. The unreported metric is frequently where the customer pain actually sits, precisely because no one is being held accountable for it.
  • What's genuinely hard here? Easy things get competed away within months. Difficulty — messy integration, regulated environments, verification, unglamorous edge cases — deters entrants and makes competence durable.

Turning intelligence into technology strategy

Most CI programmes die the same death: they become a newsletter. Well-researched, widely distributed, acted upon by nobody.

The fix is structural. Intelligence should be produced against decisions, not against competitors. Before commissioning any analysis, name the decision it will inform and the person making it.

Three decision types where competitive intelligence genuinely changes the answer:

Build-versus-differentiate. A competitor ships something you lack. The reflexive response is to match it. The better question is whether you have any evidence customers want it — including evidence the competitor themselves had. Matching a feature your rival built on a hunch is inheriting their mistake at your cost, and it's the most common way CI destroys value rather than creating it.

Positioning. Where does the market have crowded language and uncrowded reality? If every vendor claims the same thing and few deliver it, the opportunity isn't a new claim — it's proof. Publishing what you can substantiate, in a category where nobody substantiates anything, is a durable position that costs competitors real work to copy.

Pricing. Competitor pricing tells you what the market has been trained to expect, not what your thing is worth. The useful analysis isn't "what do they charge" but "what do they charge for" — the metering unit, what's bundled, where the cliff sits. A different unit of value is a strategic position; a lower number is a discount.

And one negative rule worth holding: a competitor's move is not evidence of demand. They may have less information than you, or a different customer, or the same hunch you're about to validate for them. Treat their roadmap as a hypothesis about the market, not a finding.

What AI is genuinely good for here — and what it isn't

AI has changed CI operationally, and the honest division of labour is fairly clear.

AI is good at the work humans do badly at scale: continuous monitoring across many sources, change detection, transcribing and mining sales calls for competitive mentions, summarising large volumes of buyer feedback, and spotting recurring themes across dozens of win-loss interviews that no analyst would catch manually.

AI is bad at the work that determines whether any of it matters: judgment about intent, distinguishing a routine announcement from a strategic pivot, and knowing which signal deserves a decision. One practitioner assessment puts it well — tools attempting to fully automate the analyst role produce confident-sounding nonsense.

An AI summarising unverified sources produces an authoritative-sounding document with the same error rate as its inputs, minus the caveats.

Given the citation-fabrication and fake-review data above, an automated CI pipeline with no verification layer is a machine for laundering unreliable information into confident strategy.

If you automate, automate collection and detection. Keep verification and interpretation human, and require every material claim to carry a source and a date. The pattern that works looks like: automated ingestion for speed, human validation at the point where a signal becomes a claim.

The boundary worth stating

Competitive intelligence is the analysis of publicly and legitimately available information. It is not misrepresenting yourself to obtain information, inducing someone to breach a confidentiality obligation, or acquiring trade secrets.

The line matters practically, not just ethically: intelligence obtained improperly cannot be used openly, which means it cannot inform a decision anyone can defend. If a finding can't be cited in a board paper, it has no strategic value regardless of how good it is. The simplest test is whether you'd be comfortable explaining how you obtained something to the person you obtained it from.

What to do this week

  1. Take your current competitive matrix and mark each cell by evidence source. Vendor claim, or observed behaviour? If most cells are vendor claims, you have a marketing summary, not intelligence.
  2. Pull your three closest competitors' job postings. Read them as roadmap documents. This is a one-hour exercise that routinely produces the quarter's most useful finding.
  3. Screenshot their pricing pages and diff them monthly. Cheap, automatable, and price moves are the least reversible signals a company emits.
  4. Systematise win-loss. Ten structured conversations with prospects who chose someone else will outperform any subscription you could buy.
  5. Name the decision each piece of intelligence serves. Anything that doesn't serve one is a newsletter.
  6. Add a verification standard. Every material claim gets a source, a date, and a note on whether it's observed behaviour or a company's own assertion. That single discipline separates intelligence from aggregation.

The competitive advantage in tech CI is no longer having more information. Everyone has more information. It's having information you can trust, attached to a decision someone is actually going to make.

Sources: Gartner press release, 25 June 2025, on agentic AI vendor claims; Columbia School of Nursing / The Lancet, "Fabricated citations: an audit across 2.5 million biomedical papers" (2026); "Large Language Models as 'Hidden Persuaders': Fake Product Reviews are Indistinguishable to Humans and Machines" (arXiv, 2025).

Want your competitor monitoring running continuously instead of quarterly? Get in touch.