FFrontier AEO trackerV5 across situations · September 2026
SOURCES / CITATION STUDY

Where models point you

See which websites appear in saved recommendations, then inspect the pages and questions behind them.

FOR WEBSITE OWNERS

Can models read your website?

Look up a website to see recorded access, recurring failures and the pages that made it into answers.

Looks up the saved study, not a search-engine index. A full URL is treated as its hostname; paths are not checked separately. Subdomains are included by default, so sentry.io also covers docs.sentry.io.

FROM EVIDENCE TO ACTION

Who can fix what?

All 8,821 failures · Whole-cohort analysis, independent of the explorer filters · Download full report

For website owners

Start with the exact URL. Check public reachability, stable routes, response formats and intended access rules.

A blocked tool call does not mean your website is broken.

For harness & browsing-tool owners

Verify approvals, validate tool inputs, handle supported formats and retain useful failure evidence.

An answer finishing does not mean its research succeeded.

Observed patterns and actions for both owners are shown below. Open the cited examples for supporting records. These are investigation steps, not assignments of blame.

Approval rejected5,973 67.7% of failures

What the traces show

All 5,973 occur in Cursor/Grok. The largest component is 3,394 requests without a URL, principally search requests. This is a model-runner concentration, not a pattern of websites refusing crawlers. Even public government and documentation URLs were rejected.

Website owner

No website change indicated. Your page may never have been requested.

Harness / browsing-tool owner

Configure scoped web-tool approvals; verify search and fetch with a canary before collecting answers.

Cited examples and successful counterexamples

platform.openai.com: 98 failures of this type
Example requested URL ↗
Grok 4.6 · model-speech-speech · baseline
Record 7218dff2399850dd. Same host: 115 successful operations across the cohort (not necessarily this URL).

ai.google.dev: 69 failures of this type
Example requested URL ↗
Grok 4.6 · model-speech-speech · baseline
Record b5b810fab0ba1a12. Same host: 732 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

URL safety verification failed1,123 12.7% of failures

What the traces show

The pattern spans both Codex models and both Claudes, but 960 of 1,123 events come from Claude. Hugging Face and GitHub lead the named hosts. Familiar, public hosts appearing here shows why this label must not be read as a malware verdict.

Website owner

Verify public DNS, TLS and canonical redirects. Check whether a request reached your edge.

Harness / browsing-tool owner

Expose the URL-validation reason and distinguish preflight rejection from a server response.

Cited examples and successful counterexamples

huggingface.co: 36 failures of this type
Example requested URL ↗
GPT-6 Astra · model-embeddings · baseline
Record df05c15068d9dda3. Same host: 644 successful operations across the cohort (not necessarily this URL).

github.com: 35 failures of this type
Example requested URL ↗
Claude Fable 5.1 · model-image-editing · baseline
Record 233f5b8920afb979. Same host: 703 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

Other / insufficient error detail541 6.1% of failures

What the traces show

411 of 541 come from Codex, where some saved results contain only an Internal Error header. There are also generic fetch failures from SWE and Claude. The category mixes mechanisms; rankings inside it cannot identify a shared root cause.

Website owner

Correlate the exact URL and time with edge and origin logs before changing policy.

Harness / browsing-tool owner

Retain sanitized errors, response metadata and timings so a generic failure can be diagnosed.

Cited examples and successful counterexamples

docs.n8n.io: 15 failures of this type
Example requested URL ↗
GPT-5.6 Sol · personal-self-hosted · baseline
Record 93e6ab7f095bd799. Same host: 19 successful operations across the cohort (not necessarily this URL).

statista.com: 13 failures of this type
Example requested URL ↗
GPT-6 Astra · market-research · baseline
Record e6d55c693b9a1900. Same host: 20 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

HTTP 403 — refused request268 3.0% of failures

What the traces show

211 of 268 occur in SWE. Semantic Scholar, Schwab Asset Management, SEC and Make recur. This crosses scholarly, financial and software sites; selective client/IP policy is a hypothesis, not a proven shared explanation.

Website owner

Review the precise deny rule if the resource should be public. Preserve intentional access controls.

Harness / browsing-tool owner

Record client identity and the response; do not infer a WAF or paywall from 403 alone.

Cited examples and successful counterexamples

semanticscholar.org: 12 failures of this type
Example requested URL ↗
GPT-6 Astra · research-discovery · baseline
Record e2980437e4015d77. Same host: 14 successful operations across the cohort (not necessarily this URL).

schwabassetmanagement.com: 10 failures of this type
Example requested URL ↗
SWE-1.7 · mutual-funds · baseline
Record 20dec07d9cceff66. Same host: 19 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

HTTP 404 — missing resource235 2.7% of failures

What the traces show

SWE accounts for 177 of 235; Grok accounts for 27. Cursor documentation and Zustand paths recur, alongside local-service sites. This is consistent with stale or guessed deep links, but the trace alone cannot distinguish a model-invented URL from a removed page.

Website owner

Redirect genuinely moved pages, update internal links and keep real 404s for nonexistent paths.

Harness / browsing-tool owner

Use canonical navigation or search to recover stale links; preserve the originally requested URL.

Cited examples and successful counterexamples

cursor.com: 27 failures of this type
Example requested URL ↗
Grok 4.6 · coding-delegated · practitioner-choice
Record 523574580de512a7. Same host: 104 successful operations across the cohort (not necessarily this URL).

zustand.docs.pmnd.rs: 12 failures of this type
Example requested URL ↗
GPT-6 Astra · state-management · baseline
Record eea71b737ece8f5b. Same host: 12 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

Timeout212 2.4% of failures

What the traces show

All seven models have timeouts. Microsoft.com accounts for 42 of 212, while failures also affect investor pages, workspace pages and GitHub. Cross-model presence argues against treating every timeout as one CLI approval defect. It does not establish a site-wide outage.

Website owner

Measure response times and inspect slow origin work or redirect loops for the affected URL.

Harness / browsing-tool owner

Record the timeout threshold and elapsed time. Use bounded retries, not silent unlimited retries.

Cited examples and successful counterexamples

microsoft.com: 42 failures of this type
Example requested URL ↗
Claude Fable 5.1 · team-chat · baseline
Record e79a5588858223e7. Same host: 99 successful operations across the cohort (not necessarily this URL).

ir.kuaishou.com: 11 failures of this type
Example requested URL ↗
Claude Fable 5.1 · model-image-video · baseline
Record 3652a3f579970246. Same host: 0 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

Connection / DNS failure209 2.4% of failures

What the traces show

205 of 209 occur in Claude; the other four are Muse search transport failures. Google documentation, local real-estate sites and HumanSignal recur. That distribution may reflect transport or logging differences as well as site reliability.

Website owner

Check DNS, TLS and connection-reset logs from independent networks.

Harness / browsing-tool owner

Distinguish DNS, TLS, socket and provider transport errors; capture timing and whether any response arrived.

Cited examples and successful counterexamples

docs.cloud.google.com: 12 failures of this type
Example requested URL ↗
Claude Fable 5.1 · model-image-video · baseline
Record 2c6b6587c3b83cba. Same host: 270 successful operations across the cohort (not necessarily this URL).

homesofsacramentocounty.com: 12 failures of this type
Example requested URL ↗
Claude Fable 5.1 · residential-brokers · baseline
Record 0bcc96dcbc6526ce. Same host: 0 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

Rate limiting105 1.2% of failures

What the traces show

81 of 105 occur in Muse; 84 have no named target host. Much of the signal is search-provider capacity, not a website throttling the model. Named targets include Devin, Redfin and Fiscal, which deserve separate investigation.

Website owner

For your own 429s, document limits and return useful retry guidance. A search-provider quota is outside your control.

Harness / browsing-tool owner

Reduce concurrency, respect backoff, cache repeated research and identify which service issued the limit.

Cited examples and successful counterexamples

devin.ai: 7 failures of this type
Example requested URL ↗
SWE-1.7 · coding-delegated · baseline
Record 7e716da60646f690. Same host: 30 successful operations across the cohort (not necessarily this URL).

redfin.com: 4 failures of this type
Example requested URL ↗
SWE-1.7 · residential-brokers · baseline
Record 313eaa6ef8b7c812. Same host: 27 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

Unsupported content type93 1.1% of failures

What the traces show

All 93 report text/markdown, and all are in Codex: 76 Astra, 17 Sol. Sentry contributes 30, Neon 14 and Prisma nine. This concentration suggests reader/representation compatibility rather than a general failure to fetch documentation.

Website owner

Respect requested representations, keep HTML available and separate cache variants. Do not assume an AI user agent supports Markdown.

Harness / browsing-tool owner

Support the received format or explicitly negotiate one the reader can parse. Record request and response headers.

Cited examples and successful counterexamples

docs.sentry.io: 30 failures of this type
Example requested URL ↗
GPT-5.6 Sol · observability · baseline
Record 3c15812891499bb6. Same host: 17 successful operations across the cohort (not necessarily this URL).

neon.com: 14 failures of this type
Example requested URL ↗
GPT-6 Astra · managed-databases · baseline
Record 6f60058dcbcfc200. Same host: 36 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

Invalid tool arguments35 0.4% of failures

What the traces show

27 of 35 occur in Opus and eight in SWE. Twenty-nine lack a URL. The underlying audit found malformed search JSON and unsupported fetch fields such as offset and limit: these requests can fail before any network access.

Website owner

No website change indicated: invalid research-tool arguments can fail before a request is sent.

Harness / browsing-tool owner

Validate the tool schema and retry with supported arguments. Keep these errors separate from site availability.

Cited examples and successful counterexamples

amplitude.com: 1 failures of this type
Example requested URL ↗
SWE-1.7 · product-analytics · baseline
Record 93322ade20c22bb4. Same host: 71 successful operations across the cohort (not necessarily this URL).

daily.co: 1 failures of this type
Example requested URL ↗
SWE-1.7 · realtime-media · baseline
Record 5f0db8cb84d3fb7c. Same host: 50 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

HTTP 401 — authentication required7 0.1% of failures

What the traces show

All seven occur in SWE: six target Hugging Face and one api.openai.com. This is a small cluster of potentially gated resources and an API endpoint, not broad evidence that public documentation is inaccessible.

Website owner

Keep intended authentication; offer public descriptive pages and clear authorized-access instructions.

Harness / browsing-tool owner

Use public documentation as the research source. Do not treat credentialed API endpoints as ordinary pages.

Cited examples and successful counterexamples

huggingface.co: 6 failures of this type
Example requested URL ↗
SWE-1.7 · model-text-image · baseline
Record 0de6575e4de9df3e. Same host: 644 successful operations across the cohort (not necessarily this URL).

api.openai.com: 1 failures of this type
Example requested URL ↗
SWE-1.7 · model-vision · comparative-selection
Record 9c4471cc8316ab66. Same host: 0 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

Response too large6 0.1% of failures

What the traces show

All six occur in Codex. Cohere reference pages account for three and Claude API documentation for one; two have no resolved host. The small sample points toward documentation payload limits, not a general site availability problem.

Website owner

Offer focused, bounded documentation pages and avoid unnecessarily large application payloads.

Harness / browsing-tool owner

Report the reader’s byte limit and support scoped retrieval where available.

Cited examples and successful counterexamples

docs.cohere.com: 3 failures of this type
Example requested URL ↗
GPT-6 Astra · model-embeddings · tradeoff-led
Record c11e92d3bb8c8aa5. Same host: 176 successful operations across the cohort (not necessarily this URL).

platform.claude.com: 1 failures of this type
Example requested URL ↗
GPT-6 Astra · model-chat · direct-fit
Record 5fdc4e9b1f9f5e55. Same host: 434 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

HTTP 500 — server error4 0.0% of failures

What the traces show

All four occur in SWE; three point to Technavio and one to an insurance site. These are repeated observations at two hosts, not four independent incidents. The origin versus intermediary responsible is not identified.

Website owner

Check application and gateway logs, dependencies and route health for the recorded time.

Harness / browsing-tool owner

Retain status and error details, distinguish intermediary errors and use conservative retries.

Cited examples and successful counterexamples

statistics.technavio.org: 3 failures of this type
Example requested URL ↗
SWE-1.7 · market-research · practitioner-choice
Record e85f408b90c2468c. Same host: 0 successful operations across the cohort (not necessarily this URL).

itbusinessinsurance.com: 1 failures of this type
Example requested URL ↗
SWE-1.7 · insurance-business · baseline
Record 5ba78ba8395dfc0e. Same host: 0 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

Access challenge in returned content4 0.0% of failures

What the traces show

All four detections occur in SWE: three on an SF Chronicle article and one on PyPI. The reader got a response, but it asked for JavaScript or human verification. This illustrates why HTTP success and usable research content differ.

Website owner

Where automated public reading is intended, provide a supported static path or documented API.

Harness / browsing-tool owner

Detect challenge content rather than counting any returned body as a successful read. Respect access policy.

Cited examples and successful counterexamples

sfchronicle.com: 3 failures of this type
Example requested URL ↗
SWE-1.7 · insurance-home · baseline
Record 2b40ec7727a27221. Same host: 0 successful operations across the cohort (not necessarily this URL).

pypi.org: 1 failures of this type
Example requested URL ↗
SWE-1.7 · ai-sdk · baseline
Record db0561c6548f25ea. Same host: 4 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

HTTP 402 — payment-related status4 0.0% of failures

What the traces show

All four occur in SWE: three on docs.sweep.dev/pricing and one on Investopedia. Four status observations do not establish four paywalls; 402 has no standardized payment workflow and site implementations vary.

Website owner

Inspect the actual response and configuration; publish clear public pricing/access information where intended.

Harness / browsing-tool owner

Retain the response. Do not infer a standardized paywall workflow from status 402.

Cited examples and successful counterexamples

docs.sweep.dev: 3 failures of this type
Example requested URL ↗
SWE-1.7 · coding-delegated · baseline
Record a231e48eab99cf3f. Same host: 0 successful operations across the cohort (not necessarily this URL).

investopedia.com: 1 failures of this type
Example requested URL ↗
SWE-1.7 · robo-advisers · comparative-selection
Record 2113f6a12e60c74f. Same host: 0 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

HTTP 503 — unavailable service2 0.0% of failures

What the traces show

Both observations occur in SWE at prompts-gpt.com, on the homepage and pricing page. This is one-host evidence with too little coverage to infer a broader trend.

Website owner

Check maintenance, origin health and upstream availability; provide retry guidance for temporary outages.

Harness / browsing-tool owner

Use bounded retry/backoff and retain the initial failure rather than replacing it with the retry outcome.

Cited examples and successful counterexamples

prompts-gpt.com: 2 failures of this type
Example requested URL ↗
SWE-1.7 · aeo · comparative-selection
Record 71416fce22c6fe61. Same host: 0 successful operations across the cohort (not necessarily this URL).

Inspect matching attempts →

Scope, interpretation limits and HTTP references

Historical September 2026 research traces: 59,764 recorded operations across 5,586 answers and 6,077 retained runs. This report analyzes all 8,821 explicit failures, independently of the interactive explorer filters. Counts include retries and differ across tool transports; they are not website uptime measurements or a fair model reliability leaderboard.

Evidence and interpretation limits

Each row links to matching saved observations and an example requested page. The live page link does not prove its historical status. Host counts pool prompt frames and retries; repeated failures are not independent incidents. Missing target URLs remain unattributed. Different tools expose different error detail. The 541 “other” failures remain unresolved rather than receiving invented diagnoses; another 192 unknown outcomes and 10 empty results sit outside the failure denominator.

HTTP references: Accept negotiation, Vary and representation selection, HTTP client-error semantics, and 429 and Retry-After. Recommendations above are diagnostic proposals based on the trace patterns, not confirmed root causes for each site.

What can you learn from these sources?

Choose a question, then open a website to inspect the actual cited pages. A citation shows that a page appeared in an answer; it does not prove that it caused a recommendation.

For your own site, investigate whether the cited pages explain the product’s use case, eligibility and limitations clearly. Explore website access lessons ↓

More filters & counting options

Debug retrieval operations ↗ · Request bulk access ↗