agentspeed.
What's new

What's new in AgentSpeed

The latest shipments: observability, journey canaries, alerting, and the surfaces around them.

Releases
Changed

The sidebar now asks your questions, not ours

  • Dashboard navigation is regrouped around what you are trying to find out: “Can agents find & understand us?” (watches, AI visibility, benchmarks, AI traffic), “Can they transact?” (transactions on testnet, journeys), and “What changed?” (projects, alerts, status pages). Same destinations, same URLs — nothing was added, removed, or gated; only the grouping changed.
  • The former “Agent telemetry” group is gone as a heading. Those screens now live under “What changed?”, where they support the website workflow rather than standing as a separate product line.
Changed

One workflow: scan, monitor, investigate, certify

  • AgentSpeed now presents itself as one product with four stages instead of a bundle of tools: scan a site free, monitor it with watches, investigate with one-time evidence reports (Deep Scan, Transaction Audit, AI Market Report), and certify the result. The homepage, pricing page, docs, and dashboard navigation all follow that order now.
  • The top navigation got smaller. Agent monitoring (/observability), the live demo, and the free-tools index left the top bar — every one of those pages still exists at the same URL and is linked from the footer; nothing was removed or redirected. Can agents buy? and the leaderboard took their places.
  • The telemetry line for teams building agents — ingest, runs, journeys, alerts, status pages — is unchanged and stays fully supported. It is now labeled as its own line rather than blended into the site-readiness story, and the dashboard sidebar groups it under “Agent telemetry”.
  • Honesty fixes shipped alongside: the six one-time and subscription checkout pages now say up front when checkout is not available on a deployment instead of collecting your URL and email first; the AI-traffic page no longer claims a metering arrangement that did not exist; and plan cards now list the AI-visibility and evaluations allowances that were always purchasable but never shown.
Improved

Rubric r2026.10.4: the text-density finding now tells you where your bytes went

  • The text-density check reports how much of your page is readable text. When it failed, it blamed “heavy ad or navigation markup”. For most modern sites that is the wrong culprit: the weight is the state your framework embeds in a script tag so the browser can start up, which on many pages is a second copy of the text already in your HTML.
  • The finding now measures and reports the split, so it says how much of your document is inline script and how much is markup, and the suggested fix addresses whichever one you actually have.
  • We checked whether the thresholds themselves were wrong, because our own site fails this check. Across twenty live sites they hold up: hand-built, document-first sites pass comfortably, application-style sites do not, and 30% of the sample passes. Lowering the bar would have raised our own score and helped almost nobody else, so the bar has not moved and we still fail it.
  • Two fixes ship alongside. The same page could previously score differently depending on which of our scanners read it, because three copies of “what text can an agent see” disagreed about a few HTML elements; there is now one answer. And page weight can be failed again: documents too large for us to download were being recorded as “not measured” and scored as if they were fine, which meant the heaviest pages on the web were the ones we could not fail.
  • Sites serving an oversized document score lower, and a small number of pages score slightly lower on text density. Nothing scores higher. Because the rubric version changed, that is a remeasurement rather than a change to your site, so neither drift monitoring nor score-change email will fire on it.
Fixed

Rubric r2026.10.3: a page we were never served is not a page we can grade

  • When a site answered our scanner with an error or a block instead of the page, we measured the error. Every content check ran against those bytes and reported on them as though they were the site.
  • The result was upside down. Across our stored scans, the three best text-quality scores belonged to a 404 and two blocked pages, at 90.0%, 77.6% and 40.2%, because an “Access Denied” page is almost entirely text and carries no markup, no cookie banner and no scripts. Sites that served us a real page scored worse than sites that served us nothing.
  • A non-success response now means the page was not measured. Checks that read your robots.txt, your sitemap and your /.well-known files still report, because those are fetched separately, and the reachability check still fails, because that is where a refusal belongs. Reports now say which status we got and that nothing was measured, instead of publishing a grade for a page we never saw.
  • Two related corrections ship with it. The text-density check now needs at least 2 KB of document before it will judge, because below that there is no page weight to crowd out your text. And a well-known file that answered with a block or a rate limit is no longer reported as missing: we say we could not tell, rather than telling you to publish a file you may already serve.
  • Sites that refuse our crawler, or whose homepage errors, score lower under r2026.10.3. Sites that serve a page are unaffected. Because the rubric version changed, that is a remeasurement rather than a change to your site, so neither drift monitoring nor score-change email will fire on it.
Fixed

Rubric r2026.10.2: writing about filters is not the same as offering them

  • The filter-addressability check asks whether an agent can reach a filtered view by URL. It failed a page when filter wording appeared and no filter control was present in the HTML, on the theory that the controls must be rendering only after scripts.
  • That misread an entire class of page. Its wording list mixed labels a page shows because it is drawing a control (“Sort by”, “Filter by”) with words any page might use in a sentence (“filters”, “refine”). A documentation page, a blog post, or a marketing page about faceted search would take a critical failure for a list it never had.
  • A failure now requires the stronger wording. A passing mention on its own is reported as not applicable, with a line explaining why we declined to judge. Category pages are unaffected: they say “Sort by”, or they ship controls, and either route still reaches a verdict.
  • We found this on our own homepage, which said “Filters” four times while describing this very check and offered no filtering at all. Affected pages score higher under r2026.10.2, and because the rubric version changed that is a remeasurement rather than a change to your site.
Fixed

Rubric r2026.10.1: a declared price has to be a price

  • The price-coherence check compares the price your JSON-LD declares against the price your page shows. It read the number out of the declared field by stripping everything that was not a digit, which is the right approach for a price found in ordinary page text and the wrong one for a structured field.
  • The effect: a declared price of “79.00 Save 20% today” read as 79, matched a visible $79, and passed. The check was reporting that your two channels agreed, on the strength of a number it had pulled out of something that was not a price.
  • A declared price may now carry a currency symbol or code and nothing else alphabetic. Plain long values like 1299.00 are unaffected, which was the regression worth guarding against.
  • Only sites whose price field does not contain a price move, and they score lower under r2026.10.1. Because the rubric version changed, that is a remeasurement rather than a change to your site, so neither drift monitoring nor score-change email will fire on it.
New

New checks: can an agent actually work your controls?

  • Rubric 2026.10.0 adds two checks that grade something the rubric never graded before. Everything until now measured what a page says. These measure whether an agent driving the page can operate it.
  • "Filters reachable by URL" asks whether a filtered or sorted view has its own address. Across 300 tasks on 136 live websites, filter and sorting errors caused 57.7% of the strongest agent’s failures, ahead of every other cause. If an agent can request ?sort=price_asc it succeeds on the first try; if it has to work a slider blind, it stalls.
  • "Primary action is clickable" asks whether your call to action is a real link or button rather than words painted into an image. In controlled testing, a call to action baked into an image was clicked 0 times out of 10 by every model tried, against 7 out of 10 for the same action as a real element.
  • Both are marked emerging and graded asymmetrically: they pass on the first working path, skip entirely on pages that make no such offer, and fail only when a page asks for something and provides nothing operable to do it with. Your score may move without your site changing, which is what a rubric version bump means, and the previous rubric stays readable so old scans still render.
New

AgentSpeed Certification: a verifiable, annual attestation

  • Certify your site against the published rubric: Gold (90+), Silver (80+) or Bronze (70+). Every certificate names the rubric version that produced its score, carries a public verification URL anyone can check, and runs a full year, at $199 per site.
  • The year includes weekly re-scans of the certified domain. If it falls well below your certified threshold on consecutive scans you are warned first, and a certificate that no longer holds is revoked. The verification page always tells the truth, which is exactly what makes a valid one worth displaying.
  • A new "certified" badge variant shows your tier while the certificate is valid and automatically falls back to the ordinary score badge when it is not. Scores under 70 cannot buy certification at any price.
New

The public leaderboards now have real data

  • Free scans now feed the public score API and every category leaderboard. Previously only the paid Deep Scan did, so /api/score returned 404 for a domain you had just scanned and all sixteen leaderboard pages read "not enough scan data".
  • We are also scanning a curated corpus of well-known sites across all fourteen categories, a few an hour over several days, so you can see where you stand against comparable sites rather than against an empty page.
  • Any site whose robots.txt disallows AgentSpeedBot is excluded before we fetch it, not hidden afterwards. Removal on request is standing policy.
Improved

Watch limits now differ by plan

  • Each plan now runs a set number of watches at once: 3 on Starter, 10 on Growth, 25 on Pro. Previously every paid plan was unlimited, which meant Starter and Pro bought exactly the same monitoring.
  • Only running watches count. Pausing one frees a slot immediately and keeps its entire snapshot history. You never have to delete a watch to try a different URL.
  • If you are currently over your plan’s limit, the newest watches are paused rather than deleted, oldest kept. Nothing is lost, and moving up a plan resumes them.
  • The hourly sweep now stops on a time budget instead of a fixed count, so fast checks no longer wait behind slow ones.
Fixed

Rubric r2026.07.4: no penalty for schema.org types we don’t check

  • schema.org has hundreds of types and our validator implements twenty. If your page used one of the others (SoftwareApplication, Course, JobPosting, Book), we were reporting it as a warning worth 5 points off your structured-data score.
  • That docked you for a gap in our coverage, not for anything wrong with your markup. Those blocks are now informational and excluded from scoring, and the report says plainly how many were skipped and why.
  • Affected sites score higher under r2026.07.4. Because the rubric version changed, that is a remeasurement rather than a change to your site, so neither drift watches nor score-change emails will fire on it.
Fixed

Rubric r2026.07.3: JSON-LD @graph scoring fix

  • If your structured data uses @graph (the standard way to declare several entities on one page), we were wrongly reporting a missing @context on every node inside it, and docking your structured-data score for it.
  • @context scopes the whole JSON-LD document, so nodes inside @graph correctly leave it out. Our validator was checking them in isolation. Fixed in the scanner and in the public structured-data validator.
  • Sites using @graph score higher under r2026.07.3. Because the rubric version changed, this is a remeasurement rather than a change to your site, so score-drift watches will not alert on it.
New

Score drift monitoring

  • Watch your agent-readiness score itself, not just one check at a time. We re-score your site on your schedule and alert you when the composite drops or a check regresses.
  • Two things never fire an alert: a score change caused by a new rubric version, because that is a different instrument rather than a change to your site, and movement explained only by timing-derived checks, which shift run to run on a site nobody touched.
  • The score we watch is the same one the public scan page and your badge show, so a watch and a badge can never disagree.
New

Competitor benchmarking

  • Track competitors alongside your own site and see who is actually ahead, re-scored daily.
  • Beyond the score column: the checks a competitor passes that you do not, ranked by how many rivals are already ahead, each with the fix attached.
  • Starter tracks 3 competitors, Growth 10, Pro 25.
New

Plan upgrades and self-serve billing

  • Upgrade, switch plans, update your card, download invoices, and cancel without emailing anyone.
  • Plan changes are prorated automatically.
New

Deep Scan: a one-time $29 audit

  • The full picture once, without a subscription: all 30 checks, every finding kept with step-by-step remediation.
  • A PDF report lands in your inbox when the scan finishes, and stays re-downloadable.
  • No account required: pay, and the audit runs.
Improved

Blocking AI training crawlers no longer hurts your score

  • The robots.txt check now scores only answer-time agents: the ones reading your site to answer a user right now, like ChatGPT-User, Claude-User and PerplexityBot.
  • Blocking training-data crawlers such as GPTBot or CCBot is a content-licensing decision, not an agent-readiness defect, so it is now reported for information only.
  • Scores are recalculated under rubric r2026.07.2.
Improved

No more dead buttons

  • Every dashboard control whose backing service is not configured now says so plainly instead of failing after you click it.
  • When the service goes live, the control lights up on its own.
New

AI citation tracking

  • Watch whether an AI assistant actually cites your site when answering a question you care about, and get alerted when that changes.
  • Probes Claude with live web search today. The trend over time is the signal, not any single answer.
New

Outcome and prompt-version tags on runs

  • Tag any run with a business outcome and the prompt version that produced it.
  • Compare success rate, latency and cost across prompt versions to see whether a change actually helped.
  • Available in both SDKs and the ingest API.
New

Run a journey canary from your deploy pipeline

  • POST to /api/v1/canaries/run with your API key to execute a journey on demand.
  • The pass/fail result comes back synchronously, enough to fail a deploy before it reaches users.
New

Read API and CSV export for runs

  • Pull your data back out: GET /api/v1/runs with filters and pagination.
  • Or export the runs table straight to CSV from the dashboard.
New

Journey canaries

  • Define a critical user journey as ordered, plain-language steps against any URL.
  • On a schedule we probe the live site, verify each step against what it serves, and capture each run as a full Run + span trace, with per-step pass/fail, timing, and tokens.
  • On failure, alerts fire to your channels with a deep link to the failing step’s trace.
  • Bot protection is detected and reported as “blocked,” never bypassed.
Improved

Responsive layout + accessible mobile navigation

  • The marketing site and product mockup now reflow cleanly down to mobile widths.
  • A new mobile menu: hamburger → slide-in panel with focus trap, Escape-to-close, and proper ARIA.
  • Lifted secondary-text contrast toward WCAG AA.
New

Weekly digest email

  • A Monday summary per project: success rate, p95 latency, cost, and cost-per-successful-run, week over week.
  • Highlights the top rising error type and the worst-performing tool, in plain language.
  • Per-user opt-out; idempotent so a re-run never double-sends.
New

Alert rules + channels

  • Fire on failure rate, p95 latency, error count, cost, or no-data, per agent or per project.
  • Deliver to email, Slack, or webhook, with a 30-minute cooldown to prevent alert storms.
  • Incidents open and resolve automatically as metrics breach and recover.
New

Public status pages

  • Publish a hosted status page with 30-day uptime bars per agent.
  • A public, always-on reliability signal you can share with customers.
What's new: AgentSpeed changelog · AgentSpeed