← All guides · Published November 14, 2026 · Reviewed November 14, 2026 · By the Ingenious AI team · 6 min read
Incident: What Happens When a Block Rule Breaks Your Link Previews?
The short answer: this: your shared links silently render as bare URLs — no image, no headline — in Facebook and WhatsApp, and you find out from a customer's screenshot weeks later, after the share-driven orders have already sagged. The cause is almost always a broad rule that swept up facebookexternalhit, the link-preview fetcher. The diagnosis is two checks (share-test a page; grep the block log for the fetcher's name) and the permanent fix is the policy allowlist, which subtracts the fetcher from any future blocked set by construction.
Build my block kit — free · Test a policy in the simulator first
What is a preview-breaking incident?
a crawler-policy failure in which the link-preview fetcher (facebookexternalhit) is refused, causing shared links to render without their Open Graph image, title and description. Its signature is silence: no error anywhere, only declining share performance and, eventually, a screenshot.
The symptom chain, in order
Week one: nothing. The block deploys cleanly, dashboards are green, the policy's author is pleased. Week two: a vague softness in traffic from social — attributable to anything, attributed to nothing. Week three: a measurable sag in share-driven orders or enquiries, and finally the screenshot: a loyal customer showing the owner what the weekend's link looked like in the family chat. A grey box. Every share for three weeks had been a grey box, each one a small act of anti-marketing performed at the exact moment of recommendation.
The chain's cruelty is its silence. Blocking a crawler produces logs; breaking a feature your customers use inside other companies' apps produces no signal in any system you own — except the block log, which contained the diagnosis from hour one, unread.
The two-line diagnosis
Check one: paste your own link into a WhatsApp chat. Grey card: the fetcher is being refused somewhere in the chain. Check two: search the block log — plugin table, Worker events — for facebookexternalhit. Refusals present: the case is closed, and the timestamps tell you the day it started, which will match a policy change with embarrassing precision. Refusals absent but the card still grey: the refusal lives in another layer — a CDN rule, a host firewall — and the enforcement-gap audit's layer-by-layer comparison is the tool that walks the chain.
Either way, the incident's duration was never a diagnostic problem. It was a monitoring problem: the log knew, and nobody's routine included reading the allow side of it. The fix therefore has two parts — the rule, and the routine.
The fix, and the permanent fix
Immediate: remove or narrow the rule that catches the fetcher — usually a family substring ('meta'), a scorched-earth import, or a wildcard decision made in a database view nobody read to the class column. Regenerate and verify the kit; share-test again; watch previews recover on the next shares (old shares' cards may stay cached as they were).
Permanent: put facebookexternalhit on the policy allowlist. From that generation onward, the generator subtracts the fetcher from any blocked set the policy would otherwise produce — including the blocked sets of future editors, future imports and future scares. The allowlist converts a lesson one person learned into a property of the system, which is the only form in which lessons survive staff changes.
The routine that prevents the sequel
Two habits, both cheap. First, the refusal-list read: before any policy ships, its impact preview is read specifically for dependencies — who are we about to refuse that we need? — with the preview fetcher as the standing first check. Second, the post-deploy share test: after every deploy, one product page and one article pasted into a chat, cards eyeballed, thirty seconds. The incident class dies when either habit lives; with both, it cannot reach week two.
File the incident like the others in this series: dates, the rule that did it, the refusal counts from the log (they quantify exactly how many previews were stillborn), and the allowlist entry that closes the class. The file is morbid reading for the next person tempted by a family rule — which is its purpose.
Five actions to take today
- Share-test today. Paste your most-shared page into WhatsApp; a grey card plus facebookexternalhit refusals in the log is the whole diagnosis.
- Find the sweeping rule. Family substrings and imported lists are the usual carriers; narrow the decision to named agents.
- Allowlist the fetcher permanently. The generator-level subtraction survives future editors, imports and scares.
- Regenerate, verify, re-test. New kit, hash check, and a fresh share test on real pages before you call it fixed.
- Adopt the two habits. Refusal-list reads before deploys and share tests after them — the routine is the actual fix.
Worked example: the bakery, reconstructed from its log
Crumb Culture's incident — told in the facebookexternalhit guide — reconstructs cleanly from its block log, which is why it anchors this incident format. A forum block list, adopted during a weekend tidy-up, carried a 'Meta crawlers' heading its author had never examined; beneath it sat the preview fetcher's name. From the Monday deploy, the log recorded facebookexternalhit refusals: 1,940 across three weeks, each one a preview that never rendered.
Share-driven orders sagged an estimated 28% before a customer's screenshot closed the loop that monitoring had left open. The fix took an afternoon: the list replaced by a graded AgentBlock policy, the fetcher allowlisted, the kit verified, previews rendering the same day.
The bakery's incident note, filed with the refusal counts, ends with the sentence its owner now gives other businesses verbatim: 'The log knew on day one. Read the log.'
Frequently asked questions
How can I tell if my previews are broken right now?
Two checks, two minutes: share one of your pages in WhatsApp and look at the card; then search your block log for facebookexternalhit refusals. Broken card plus refusals is conclusive. A healthy card means the fetcher is getting through on that page — repeat for your most-shared few.
Why didn't anything alert us?
Because nothing errored. Your site served refusals correctly; the preview failure happened inside Facebook and WhatsApp's rendering, which reports nothing back to you. The only instrument that saw it was your own block log — hence the routine of reading refusal lists and share-testing after deploys.
Do old shares recover once the fetcher is allowed again?
New shares render correctly immediately. Cards already rendered and cached by the platforms may persist in their broken form for a while, which is one more reason the incident's true cost is measured in the weeks before diagnosis, not the afternoon of the fix.
Can rate caps break previews too?
Yes, in a burst: a viral share generates preview fetches faster than a tight per-minute cap allows, and late sharers get the grey card. Size the utility agent's treatment as allow-plus-allowlist rather than a budget — its traffic is your distribution, not a load problem.
What if the log shows no refusals but cards are still grey?
Then the refusal lives outside the kit — a CDN or host rule, or stale Open Graph data on the page itself. Walk the layers with the enforcement-gap audit's comparison, and check the page's OG tags directly. The incident class is 'previews broken', and the kit is only its most common address.
Related guides
- How to Block AI Scrapers and Bots from Your Website
- Which AI Crawlers Should You Allow?
- How Do You Block AI Bots on WordPress, Cloudflare, or Plain JS?
- Pre-deploy breakage simulator — test a policy before it ships
- Impact preview — your server log against a draft policy
- What Can Robots.txt Actually Do About AI Crawlers?
- How Do You Design an Allowlist for AI Agents?
- What Are the Most Common AI Blocklist Design Mistakes?
- Per-Agent Policy or Block All AI Bots: Which Is Right?
- How Do You Test a Block Policy Before Deploying It?
- Block evidence — parse a log, build an evidence pack
- The dated agent database — 25 entries, graded