SourcrLab Research II · August 2026
The Citation Surface Thesis
If buyers increasingly ask AI which software to choose, what information does AI actually have to work with?
This thesis also started in recruitment technology. We kept seeing the same problem: vendors publish plenty of content, but the information that actually helps compare products is often missing, vague or buried. AI does not fix that. It simply becomes another layer trying to make sense of the same market.
But the mechanism is broader than recruitment software. The same question applies to almost any software market: when an AI system has to recommend a CRM, accounting platform, developer tool, security product or marketing platform, which pieces of information are clear and useful enough to make it into the answer?
Our thesis: as software discovery moves into AI answers, clear, original and sourceable evidence becomes more valuable than generic product copy.
1. A page is no longer the only useful unit
A human may read a whole article. An AI system may only need one price, one limitation, one integration check or one comparison, then combine that with information from other sources. The page still matters, but the useful pieces inside the page matter too.
That is what we mean by a citation surface: a piece of information that is clear enough, specific enough and well-supported enough to be useful on its own.
2. Generic language gives AI very little to compare
'AI-powered', 'all-in-one', 'built for modern teams' and 'work smarter' may sound good, but they do not reduce much uncertainty. If ten vendors say roughly the same thing, they become hard to distinguish for humans and machines alike.
Weak: “Acme is an AI-powered platform helping modern teams work better.”
Useful: “Acme is primarily a CRM for small service businesses, includes email sequencing and pipeline automation, and does not publish a self-service starting price.”
The second example is useful because it removes uncertainty, not because it contains more words.
3. Being online is only step one
We separate six stages because 'AI visibility' is too vague as one score. A source can be reachable but never retrieved, retrieved but not selected, cited but not used meaningfully, or used without the brand being visible.
Available
Can the system access the information at all?
Retrieved
Does the page surface for this question or a related sub-question?
Selected
Does the system choose your source over competing sources?
Used
Does your information actually make it into the answer?
Attributed
Is your brand or source visibly credited?
Influential
Does your information change the shortlist or recommendation?
4. Evidence is harder to copy than prose
Anyone can rewrite a category article. Anyone can ask an LLM for another guide. What is harder to copy is the work underneath a claim: checking a price, testing a workflow, verifying an integration, recording a product change, collecting user evidence or comparing hundreds of products under one method.
That is why we think the long-term moat is not a GEO trick. It is evidence that somebody else has to do real work to reproduce.
5. Sourceable beats impressive-sounding
'Flexible pricing for teams of every size' sounds polished but tells a buyer little. 'No public starting price found; sales contact required; checked on 23 August 2026' is less glamorous and far more useful. The same applies to integrations, limitations, implementation and target audience.
6. Facts help. Decision rules help more.
Knowing that a product has a feature is useful. Knowing when that feature should change a buying decision is more useful. A CRM having advanced automation matters differently to a five-person team with a simple process than to a 200-person sales organisation with complex routing and governance.
This is the Decision Layer: connect product facts to the conditions under which they matter.
7. Why recruitment tech is our evidence base
SourcrLab does not claim to have measured every software market. Our first-hand research base is recruitment technology, where we can observe product overlap, pricing opacity, integrations, AI claims and buyer decisions in detail. That gives us somewhere real to test the wider thesis.
If Citation Surface is a useful general model, the same mechanisms should show up in other software categories too. That is a research question, not something we need to pretend is already proven.
8. What this means for software vendors
If a vendor wants to be represented accurately by AI systems, the answer is not to stuff pages with AI keywords. Make the product easy to understand. State the primary category. Say who it is for. Explain what it replaces and what it does not replace. Publish pricing where possible. Keep integrations current. Make limitations visible. Date material claims.
9. What this means for SourcrLab
The goal is not to make pages that ChatGPT likes. The goal is to build the strongest independent body of recruitment-tech evidence we can, then make that evidence easy for humans, search engines and AI systems to understand. The thesis is broad; the research engine remains grounded in the market we actually know.
10. Link to Research I
The Tech Noise Problem looks at the buyer side: crowded software markets create more options but not automatically better decisions. The Citation Surface Thesis looks at the information side: which evidence survives when AI systems increasingly mediate that discovery?
Read Research I: The Tech Noise Problem →Methodological boundary
This is a working thesis. We propose the mechanism across software markets, but our direct evidence base is recruitment technology. We will keep hypothesis, observation and proven result separate as the dataset grows.
Conclusion
If software discovery keeps moving from browsing pages to receiving answers, the strongest information asset will be the source with the most useful things worth reusing.