We read it for marketers and business owners, and publish only the findings you can act on.
01
We follow new papers on how AI search finds, ranks and cites sources.
02
We check each paper’s methods, results and limits.
03
Only findings with a practical use for a business make the cut.
04
We explain each one in plain English, with the evidence and a link to the paper.
Read it as text
Software follows new research on how AI search finds, ranks and cites sources, as it appears on arXiv and Crossref. Finding a paper doesn’t commit us to writing about it.
A person decides which papers are worth an article. Each one has to answer a question readers have and add something we haven’t already said. A paper that repeats a known finding can update an article we already have, or stay on file without one.
An AI reads the full paper, past the abstract to the last page, and notes its findings, methods and limits, each with the exact words that support it. Software checks every quote against the paper. If the full text isn’t available, or the notes don’t check out, the work stops there.
An AI writes a draft from those notes and our writing rules, in plain English, with a link to the paper. It uses the notes the explanation needs, and the rest stay in our files.
One AI reviewer reads the draft against the paper and asks what readers will come away believing, and whether the paper supports it. If something left out would change that impression, it has to go on the page; a limit kept only in our notes doesn’t count. A second AI reviewer reads only the article, as a reader would, and asks whether it’s worth the time. If a passage could be easier to read without losing anything, it has to be fixed. Required fixes go back to the writer, and both reviewers read the new draft. A draft that still falls short after a few tries is held.
A person reads the finished article and decides whether to publish that exact text, ask for changes or hold it. If the text changes afterwards, it needs a new decision.
Only approved text goes into the site. Once an article is live, software checks the page readers get: its text, its byline and its link to the paper.
A paper can stop at any of these steps, and there’s no quota to fill. When nothing earns a place, nothing is published.
The explainer follows a real paper, “Structured Linked Data as a Memory Layer for Agent-Orchestrated Retrieval”, to its article, Before buying more schema work, ask what reaches the retrieval pipeline. The drafts and review marks it shows along the way are illustrations.
Measuring seller influence on AI shopping recommendations
Controlled testing on GPT-4o reveals why simple win rates fail to separate strong marketing copy from weak filler—and how introducing a price penalty exposes the true strength of a preference.
Read the digestGenerative Engine Optimization
On the E-GEO benchmark, a nine-rule editing handbook delivered most of the ranking improvement over a single optimized prompt, while automated rule evolution added a smaller boost.
Read the digest19 more articles
Enabling live web search significantly alters which local providers an AI recommends. Learn why visibility audits must track search configurations and verify business identities rather than relying on simple name matches.
Standard safety filters, factual verification, and resistance to source manipulation are distinct capabilities. When evaluating AI research tools, test how defenses handle ordinary queries alongside manipulated ones.
The revised Dice Roll study retracts its fixed repetition tiers. Effective AI auditing requires a pilot to measure variance for your specific comparison, alongside a distinct design for tracking changes over time.
A link, a backed-up claim, and a true description are three very different things. To know what an AI visibility report is actually telling you, you have to read the whole answer, not just count the links.
A brand missing from a general AI category list might still show up for a specific buyer need. Learn how to separate broad visibility, realistic use cases, and diagnostic probes in your audits.
Separate query fit, observable source records, and final synthesis before commissioning a rewrite. Source logs narrow the failure mode, even when they do not expose the full retrieval pipeline.
AI visibility reports that stop at the opening prompt miss what happens next. A new theoretical paper models how citations survive or vanish across multi-turn conversations, without confusing persistence with actual buyer influence.
A small test in Tallinn shows that changing prompt language can alter which local or global accounting software suppliers ChatGPT recommends—highlighting why AI visibility tests must account for language.
A controlled simulation by Gregory Druck and Ethan Smith shows that when language models retrieve articles generated from their own prior answers, response diversity can collapse—illustrating the difference between genuine corroboration and circular repetition.
A Bali AI audit found 85.6% of listed venues never appeared in sampled answers. Here is why the study's denominator, entity matching, and ranking split matter.
An analysis of AI professional recommendations reveals the critical difference between matching answers against a preset list of names and detecting who an AI assistant actually recommends.
A study of Chinese-language generative search highlights the gap between appearing in an AI reference list and having your brand or contact details actually included in the generated response.
Discovery, citation, representation, and user action are distinct steps in an AI system's pipeline. A recent review of 45 studies shows why Generative Engine Optimization claims must be anchored to the exact stage they test.
A prelaunch evaluation of an Icelandic public-information AI flagged 35% of web-search answers and 6% of curated answers for source issues. Because the two modes were audited against entirely different criteria, those numbers cannot be treated as a head-to-head trust comparison.
An evaluation of a Glasp case study illustrates the difficulties of isolating the causal effect of content changes from broader ChatGPT platform growth. Pre-existing trends and placebo tests show why even controlled observational estimates require careful interpretation.
Before interpreting an AI brand-visibility score, you need to separate ordinary rerun variance, equivalent rephrasings, and changed buyer requirements. Each measures a fundamentally different aspect of model behavior.
Showing up once in a massive test run is not the same as showing up consistently for buyers. Learn how to check the denominator, retrieval layer, and model setup behind AI coverage claims.
Comparing Google Search and Dataset Search for AI agents reveals a classic retrieval trade-off: the specialized index delivers more usable data records, while general search covers far more queries.
Embedded markup, extracted text, and richer entity pages are distinct inputs. A controlled study shows why that distinction matters—and why better AI answers cannot be credited to formatting alone.
