Allen Review: useful map, unclear audience (73/100)

Satlas is a map-driven research tool for exploring change through AI-annotated satellite imagery.

Fact	Value
Domain	satlas.allen.ai
Category	AI-generated open geospatial data from satellite imagery
Pricing	Unknown; no pricing model or price range identified
Pages crawled	10
Crawl date	2026-10-24

Satlas is strongest when it gets visitors to the map: usability scored 91.3/100, and the sampled experience presents a clear core offer with an "Explore Map" action. Technical SEO also scored 87/100; the crawl was server-rendered across 10 sampled pages, with HTTPS and a one-hop HTTP-to-HTTPS redirect. Accessibility scored 82/100, while risk and stability scored 84/100.

The central weakness is role clarity. First impressions scored 52.2/100 and audience and messaging scored 48.1/100 because visitors still have to infer which researcher, analyst, or operational job is the priority. Design execution scored 66.4/100, and writing quality scored 72/100: the site has specific research material, but its editorial focus can be sharper. Performance is mixed, with desktop LCP at 2.819 s p75 and mobile lab LCP at 5.95 s.

Overall, Satlas is a credible open geospatial-data research surface with a strong path into exploration, but it needs clearer audience framing before its breadth becomes immediately useful to a first-time visitor.

Satlas establishes a specific category quickly, but its primary job and audience remain implicit. The hero, "Open Geospatial Data Generated by AI," identifies the data type and open-access promise, while examples such as China offshore wind farms and Amazon rainforest deforestation make the dataset tangible. The supplied crawl does not verify a named user, product metric, attributable case study, or competitive alternative that would explain why Satlas should be chosen for a particular decision. Add one role-and-job sentence beside the hero, then attach a measurable proof point such as coverage, update cadence, model accuracy, or licensing. That would sharpen the proposition without weakening its open-data character.

The homepage answers what Satlas is more effectively than who should act on it. "Open Geospatial Data Generated by AI" and the above-fold "Explore Map" action create a workable starting point, and examples of offshore wind, deforestation, and logging show what can be studied. The supplied material does not name a priority role or decision, and it does not provide adoption details for licensing, methodology, update cadence, support, or failure handling; those answers may exist outside this crawl. Put a plain-language audience-and-job sentence before the CTA, then turn one example into attributable proof with the user's role, question, and outcome. Keep licensing, methodology, and update cadence close to Map or Data.

Satlas gives visitors a strong route into exploration and practical help, with the main usability cost arriving before the first mobile interaction. The homepage pairs the offer with "Explore Map"; navigation labels such as Map, Data, and FAQ are plain; and the FAQ answers "Can I download the data?" with concrete formats including GeoJSON and GeoTIFF. Desktop and skeptical-evaluator walkthroughs completed cleanly, while the mobile run completed with struggle because mobile LCP/TTI was 5,950 ms. The strongest fix is to prioritize the hero and defer below-the-fold media so the CTA becomes usable sooner. This was a simulated crawl-based walkthrough, not live user testing, and covered 10 of a 40-page budget.

The sampled public pages have a solid baseline, but three structural fixes should precede a broader conformance pass. The homepage DOM reports 39 images with no missing or empty alt decisions, and sampled pages report no inputs without a detected label or ARIA name. PSI nevertheless reports an unnamed button and an SVG with role=img without an accessible text alternative. The captured SatlasPretrain DOM has no landmarks and no skip link, while FAQ, AI, and Data samples skip from h1 to h3. Name affected controls and informative SVGs, add a skip link and meaningful regions to the pretrain page, then normalize heading levels. Keyboard traversal, focus order, announcements, and captions remain untested in this automated artifact audit.

Satlas has a coherent visual entry point, but mobile interaction sizing is the clearest design execution gap. The sampled homepage has one H1, a 46px mobile H1 against 16px body text, and no horizontal overflow at a 390px viewport. It also reports 33 small tap targets, including 40x40px and 30x30px controls; /map reports 35, including 28x28px, 24x24px, and 20x20px controls. Bring non-exempt hit areas to at least 44x44px with 8px separation. The stylesheet inventory reports 85 colors, 15 radii, and 44 spacing values; those counts may include unused rules, so verify token drift with computed styles rather than assuming every rendered component is inconsistent. Preserve the current CLS result while sizing images explicitly.

Performance is mixed: desktop field data is near the LCP boundary, while the mobile lab path is slow to the first useful view. CrUX reports desktop LCP at 2.819 s p75, CLS 0, INP FAST, and TTFB 985 ms p75; the mobile lab reached LCP at 5.95 s versus 1.59 s on desktop. The homepage transfers 2.69 MB, /map 3.42 MB, and /superres 9.78 MB, with up to 1.65 s of mobile render-blocking savings and 499–701 KiB of estimated image savings. Prioritize the hero request, defer non-critical work, and set route budgets for /map and /superres. Mobile field data was not supplied, so the lab result is a diagnosis of conditions, not a population-wide verdict.

Satlas's strongest copy names scale, resolution, dates, and formats; its weakest surface asks users to operate the map before explaining what they can learn there. SatlasPretrain supplies "over 30 TB of satellite images" and "137 label categories", while the FAQ adds 10m/px imagery, monthly data from January 2016, GeoJSON, and GeoTIFF. By contrast, the map has no H1 or headings and begins with controls such as Application, Layout, and Time. Add an answer-first orientation paragraph before those controls. Then apply the same specificity to metadata: nine pages reuse the title "Satlas" and one shared description. Unique titles and descriptions would let each topic carry its own research intent.

The most concrete stability risk is URL handling, not measured traffic loss. A guaranteed nonexistent path returned HTTP 200 with the 1,751-byte Satlas application shell, creating soft-404 ambiguity if arbitrary invalid paths behave the same way. Return a real 404 and rerun the probe. Mobile performance is a secondary discovery-to-engagement risk: homepage lab LCP was 5.95 s, while CLS was 0 and desktop lab LCP was 1.59 s; no mobile field data was supplied. Raw/rendered parity was observed, and no migration, redirect-loop, or sitewide robots block was evidenced. GSC and analytics were unavailable, so traffic magnitude and recovery timing remain unmeasured.

Satlas has credible editorial ingredients, but the sampled publish gate misses page identity, heading structure, and claim-level linkage. The AI page cites "over 30 TB of imagery", "302 million labels", and "137 diverse categories" alongside links to SatlasPretrain, the paper, training data, and a Data Validation Report. Yet nine pages share the title "Satlas" and one description, the FAQ jumps from H1 to H3, SatlasPretrain begins with H3 before H1, and the map has no heading or orientation paragraph. Require unique metadata, valid heading order, and a supporting document at each high-stakes numeric claim. The supplied artifacts do not establish schema mismatch, dead citations, or a programmatic thin-tail breach, so those should remain open checks rather than findings.

The FAQ is a useful lightweight support surface: it is linked from primary navigation, has inDegree 8, and covers seven questions including downloads, accuracy, missing imagery, and code access. The /llms.txt and /llms-full.txt probes each returned status 200 with the 1,751-byte React application shell rather than a plain-text documentation index. Add a reviewed date, a compact answer index or search, and genuine plain-text files linking the FAQ, data documentation, validation report, and code repository.

Satlas has a strong sampled crawl baseline: all 10 inventory rows returned HTTP 200, the root HTTP URL redirects once to HTTPS, and raw/rendered word counts match at 422/422 on the homepage and 570/570 on /faq. The main technical defect is the guaranteed nonexistent URL returning HTTP 200 with the 1,751-byte application shell; make it a real 404. The mobile PSI run also reports LCP 5,950 ms, FCP 2,409.9 ms, TBT 142 ms, and CLS 0. Add matching Organization and WebSite markup and the missing headers after HTTPS validation. Sitemap coverage is unknown, not confirmed absent.

Satlas earns its 73/100 through a fast route to exploration and a clear research premise. Usability is exceptional at 91.3/100, while accessibility, performance, risk and stability, and technical SEO all score in the strong band.

The fixable weaknesses are audience definition, mobile loading, and editorial focus. Audience and messaging is 48.1/100, first impressions is 52.2/100, and the mobile lab reached LCP in 5.95 s. Writing quality is 72/100 and editorial QA is 64.1/100, suggesting that the research material needs tighter prioritization and a clearer publish gate.

Satlas is best suited to geospatial researchers, environmental analysts, and infrastructure observers seeking time-based change data. Visitors outside those roles can reach the map quickly, but may need more guidance to understand the first job the product is meant to support.

This 13-dimension review uses the supplied crawl of 10 pages from 2026-10-24, current product facts from the crawl, the site profile, and the public dimension score table. Usability evidence includes simulated cognitive walkthroughs against crawl data rather than live user testing. Performance evidence combines the supplied desktop field signal with a mobile lab run.

Dimensions 12, Decision-support surfaces, and 13, Review-content integrity, were excluded because both were marked not_applicable. GSC enrichment was not connected or accessible, and no traffic symptom was supplied; those are data gaps, not findings about the site.

Read How SiteList scores for the scoring method, dimensions, and evidence rules.