SemanticRisk findings — May 2026: AI visibility movement needs an explanation, not just a score.

Updated review language: access regression and capture contraction are now separated.

After the May report, the rolling review workflow was refined to avoid grouping different signals together. A true crawl/access regression is now treated separately from a reachable-page capture contraction. This matters because a blocked or rate-limited page, such as an ok / 200 state moving to blocked / 429, is not the same as a page that remains reachable while the amount of captured text falls sharply.

This May report keeps the original month-end findings intact, but uses the clearer language where possible: crawl/access regression for access failures and capture volatility for reachable pages with materially reduced AI-visible content.

Key takeaways

• May produced 10 reviewed findings: five top movers, three crawl/access or capture-volatility observations, and two controlled unchanged-content interpretation-drift observations.
• The strongest benchmark movers were monday.com, jacobs.com, azure.microsoft.com, roche.com, and exxonmobil.com.
• The strongest mover, monday.com, rose from 45.55 to 57.75 as the AI-visible page representation expanded and extracted product/commercial claims became more specific.
• roche.com moved in the opposite direction, falling from 46.85 to 39.45 as captured text decreased and the extracted claim set became narrower and more metadata-heavy.
• Model-disagreement findings remained excluded from the public story because the May snapshot produced no publishable disagreement findings.

AI visibility movement is not one thing.

The most useful May finding is not that several domains moved. It is that they moved for different reasons. A benchmark score can change because a page expands, because the extracted business narrative becomes clearer, because structure is interpreted differently, or because the page becomes thinner and less semantically useful to automated systems.

monday.com moved up as the AI-visible product narrative became more specific.

monday.com was the largest positive mover in the May sample, rising from 45.55 to 57.75 between May 19 and May 20. The movement appears tied to a material expansion in the AI-visible page representation. The text length estimate increased from roughly 11,781 to 16,409 characters, while normalized content length increased from about 11,860 to 16,470 characters.

The earlier scan emphasized broad positioning around monday.com as an AI Work Platform and Work OS. The later scan extracted more specific commercial and product signals, including Work Management, CRM, Dev, Campaigns, Service, monday sidekick, free and paid plans, an aggregate rating, and pre-built AI agents.

jacobs.com shifted from technical webpage signals toward business/service claims.

jacobs.com rose from 50.05 to 57.75 between May 13 and May 14. The earlier scan mostly surfaced technical webpage signals such as title tags, analytics scripts, viewport tags, CSS references, accessibility links, ARIA labels, and structured data.

The later scan extracted more business-facing claims, including Program Management, Engineering, Procurement and Construction Management, Augmented Delivery, Advanced Manufacturing, Digital Infrastructure, Energy & Power, Environmental, Health, National Security & Defense, Life Sciences, Transportation, and Water.

azure.microsoft.com moved even though captured text volume stayed broadly stable.

azure.microsoft.com rose from 50.05 to 57.75 between May 4 and May 5. Unlike some other May movers, this was not driven by a large increase in captured page text. The text estimate moved only from roughly 21,248 to 21,299 characters, while normalized content length changed slightly from about 21,405 to 21,270 characters.

The extracted claims remained centered on Azure’s cloud-computing positioning, including canonical URL, title, meta description, indexing directives, accessibility links, analytics scripts, and page metadata. The later scan produced a cleaner structural reading of the page, including navigation, language, viewport, and canonical-link signals.

roche.com moved down as the captured representation became thinner and less clear.

roche.com was the main negative mover in the May top-five set, falling from 46.85 to 39.45 between May 8 and May 9. The site remained reachable with HTTP 200 responses, so this was not a hard crawl failure. However, the captured text estimate decreased from roughly 7,403 to 6,258 characters, and normalized content length fell from about 7,420 to 6,267 characters.

The earlier scan extracted Roche’s corporate positioning along with page-structure and usability signals such as Open Graph metadata, an animated homepage headline, accessible focus-visible outlines, Gatsby site-generator signals, breadcrumb navigation, and responsive viewport configuration. The later scan still captured Roche’s core healthcare positioning, but the claim set became narrower and more metadata-heavy.

exxonmobil.com also appeared in the May top mover set.

exxonmobil.com rounded out the strongest May mover group, with a benchmark movement of approximately +6.10. It is included as a reviewed May mover, but the clearest public evidence narrative for this update comes from the more detailed examples above: monday.com, jacobs.com, azure.microsoft.com, and roche.com.

Access volatility and capture contraction both affected what AI systems could see.

May produced multiple preflight-level signals. Some were true access regressions, where a previously successful state moved to a rate-limited response. Others were better described as capture volatility: the site remained technically reachable, but the amount of AI-visible page text fell materially.

Controlled unchanged-content drift appeared again, but was not used as a public flagship example.

May produced two unchanged-content interpretation-drift observations on cr8ivtek.com and decisioncheckpoint.com. Because these are owned or controlled properties, they are useful as internal validation of the monitoring mechanism but weaker as independent public benchmark examples.

The editorial decision is to keep this signal in the report without overstating it. The May public story is strongest in the explained top movers and crawl/access regressions, while unchanged-content interpretation drift remains an internally validated monitoring lane.