Field Notes
Deep, practical Field Notes on affiliate systems, search-led content, and responsible AI. Every article shows its reasoning, limits, and next action.
A click counts only when your application checked the action and you can see the screen it changed. A model request is not permission. A closing sentence is not the result.
A mismatch between the outcome a reader expected and what the page did. A chosen installer can stay. A virus, a silent homepage switch, or a trapped back button is the pattern to retire.
A background result counts only when a later retrieve shows a terminal status and the output. A queued create is not that result. A webhook id is not the output.
Spam added through a channel you opened for readers. A real comment can stay. A promo post, a spam account, or an unmonitored upload is the pattern to retire.
An MCP result counts only when the response includes the tool list and a completed call. A server URL is not that list. A hint is not an allowlist.
A page that claims a service it does not provide. A working calculator can stay. A fake credit generator, or a merge tool that only opens ads, is the pattern to retire.
A web-search answer counts only when the response includes a search call and a citation you can open. A recited URL is not that call. A search result is not the offer.
Automated queries sent to Google, including rank-check scrapes, without express permission. A typed query can stay. A script pulling result pages is the pattern to retire.
A file search answer counts only when the index finished and the response includes a retrieval call. An uploaded file is not that index. A filename is not the passage.
Content placed on a site without permission, through a security flaw. A comment you opened can stay. An injected script, a new spam URL, or a search-only hop is the pattern to retire.
A calculated figure counts only when a code-execution result object returned it. A recap is not that object. A sandbox printout is not a write to your systems.
A redirect meant to deceive the reader or the crawler. A moved page can stay. A phone hop to a different spam domain is the pattern to retire.
The tier you asked for is not the tier you were billed for. Read the response field. Flex is a slower synchronous call, not a batch job, and Fast is not a better model.
Different content for search engines than for readers, on purpose. A paywall Google can read can stay. A keyword block that appears only for a crawler is the pattern to retire.
Keep a claim only when the citation points at a span you can open. A typed source list is not that pointer. A citation is not proof the source is true.
Content placed so search engines see it and visitors do not. A tab, a tooltip, or a screen-reader label can stay. White-on-white text, a zero-opacity block, or a hyphen-only link is the pattern to retire.
Carry the prior turn by ID or by the full item list. A new sentence is not the conversation. Stored IDs expire, and a chained turn is still billed as input.
Filling a page with keywords or numbers to move rankings. Naming the product a reader is comparing is allowed. A city list, a synonym stack, or a sentence that repeats the same phrase is the pattern to retire.
Count the full request on the model you will call. A word estimate misses tools, files, and formatting. An input count does not predict the reply.
Taking other sites' pages, feeds, or media and hosting them to move rankings. A cited spec can stay inside a page that adds judgment. A copied review, a synonym swap, or a feed with no new benefit is the pattern to retire.
A refusal is a finished response that declined the task, not a dropped connection. Check the stop field, discard partial text, and do not continue that same turn.
Links created mainly to move rankings, inbound or outbound. A paid or affiliate link can stay if it is qualified. A link bought to rank, or a page that exists only to trade links, is the pattern to retire.
Pages built to rank for similar queries, then funnel readers to a less useful stop. A page a reader would use on its own can still be a destination. A city-name swap that leads to one offer is the pattern to retire.
The agreement requires a tagged Special Link, a prominent Associates sentence, and original content. A click is not a commission. The session, cart, and 180-day ship-and-pay clocks still have to pass.
Reddit does not ban affiliate links as a class. The community rule, a plain commission sentence, and the program terms still have to pass. The Brand Affiliate tag is optional.
A moderation flag is a routing signal, not a verdict. Check it before you store, show, or act. Image-only checks miss text-only categories, and a pass is not permission to publish.
Buying a lapsed domain mainly to reuse its old search reputation for low-value content. A new people-first site on an old name is allowed. The registration does not include the old rankings.
Reasoning tokens are billed even when you never see them. Leave room in the output cap, start at low effort, and do not publish the scratch work as evidence.
Third-party pages on a host site, placed there mainly to borrow ranking signals. Your own affiliate reviews are not the violation. Since August 30, 2026, a manual action changes results outside the EEA only.
Upload once, reference by ID, and wait until the file is ready. Match the content block to the MIME type. Do not paste a long PDF or accept user-supplied file IDs.
Check the stop field before you use a model reply. Treat length, max_tokens, and MAX_TOKENS as incomplete. Do not parse truncated JSON or run a partial tool call.
Google builds snippets from page text, not only your meta description. Use robots meta or X-Robots-Tag to limit or hide them. Do not robots.txt-block a URL that still needs those rules.
Pin one embedding model and compare vectors only inside that space. Store each vector with its source text. Rank with cosine similarity. Relatedness is not a fact check.
Scaled content abuse is making many pages to manipulate Google Search, not to help a reader. AI is allowed. Unoriginal volume without added value is the problem.
Google indexes a page after it can find, crawl, and keep it. Use Search Console to see which stage failed. Requesting indexing does not guarantee inclusion.
When a review URL is gone, return HTTP 404 or 410. A 200 error page is a soft 404. Do not 301 every dead SKU to the homepage.
Give each listing slice a unique URL, sequential href links, and its own canonical. Do not point page 2 at page 1, and never put page numbers in fragments.
X allows affiliate links, but treats commissions as paid partnerships. Use the Paid Partnership disclosure, disclose clearly, and check program terms.
Use a provider batch API when work can wait. Submit JSONL with unique IDs, poll until the job ends, and match results by custom_id—not by line order.
Facebook allows affiliate links, but treats commissions as branded content. Use the Paid partnership label, disclose clearly, and check program terms.
Stream tokens over SSE so work can start before the model finishes. Accumulate deltas until a completed event, then validate. Partial JSON is not a finished tool call.
The FTC fake-review rule mostly targets sellers. Affiliate pages still need honest experience and a commission disclosure. Fake followers are banned for anyone.
Treat the context window as a hard budget. Count tokens before you send, leave room for the reply, and compact or drop history before the request overflows.
Thin affiliation is merchant copy plus tracking links with no added value. Google's spam policy allows affiliate pages that add original reviews, tests, or comparisons.
TTFB is the wait before the first HTML byte. Measure it in the field, then cut redirects, origin work, and cache misses—without treating 0.8s as a ranking score.
LinkedIn allows affiliate links in public posts, but treats commissions as a brand partnership. Use the label, disclose clearly, and never buy LinkedIn ads for the offer.
Pin production calls to a snapshot ID, not a moving alias. Log the resolved model, test upgrades on a copy of the workflow, and migrate before retirement.
Keep unfiltered listings and product pages crawlable. Block generated filter combinations unless a filtered view is a real page with its own job.
Put UTM parameters on inbound links to your pages, not on outbound affiliate hops. Keep source, medium, and campaign; leave merchant IDs alone.
TikTok allows affiliate links, but treats commissions as branded content. Use the commercial disclosure toggle, disclose clearly, and check program terms.
Put unique og:title, og:type, og:image, and og:url tags in each affiliate page head so shared links show the right preview—not a guessed logo.
Connect related reviews, comparisons, and hubs with crawlable HTML links and descriptive anchors. Every important affiliate URL needs an in-content path from another page.
Instagram allows affiliate links, but treats them as branded content. Use the Paid partnership label, disclose the commission clearly, and check program terms.
Put stable tools, instructions, and reference material first, mark a breakpoint at the last identical prefix, and measure cache reads—caching does not change answers.
Declare a small set of JSON Schema tools, execute each model call in your own code after validation, return the result, and keep write actions behind an allowlist.
Use one visible H1 that names the page's job, nest H2 and H3 by rank, write headings as section labels rather than keyword lists, and keep the outline aligned with the title.
Constrain AI workflow replies with a JSON Schema, prefer strict structured outputs over JSON mode, validate meaning, and treat refusals or truncated replies as failures.
Retry only transient AI failures, cap the attempts, wait with exponential backoff and jitter, honor Retry-After, and never retry unsafe side effects.
Write one unique, accurate summary per affiliate URL so search snippets can describe the page's real job—not a keyword list or a site-wide template.
Write each title tag as a unique, concise description of the page's job, then keep the visible H1 aligned so search, tabs, and bookmarks all name the same page.
Limit an AI agent's reach with narrow tools, scoped identities, runtime authorization, approval gates, isolation, and tests that prove denied actions stay denied.
Improve LCP by finding the page's real largest element, then shortening its request path, reducing its bytes, and removing delays before it can render.
An affiliate click usually stays with its browser and device. Cross-device credit needs an approved identity bridge, app integration, or program-specific code.
Treat every webhook as an untrusted event: authenticate the original bytes, reject replays, validate meaning, and limit what accepted events can trigger.
Improve INP by tracing the slowest real interactions, then reducing main-thread waiting, event-handler work, and the rendering needed for the next visual response.
Reserve the final space for product media, ads, embeds, fonts, and late interface elements before they load, then fix the largest unexpected movement first.
Use AVIF or WebP when a tested export meaningfully reduces image bytes without obscuring product details, and keep a dependable fallback and responsive sizes.
Use a CDN when edge caching solves measured latency, traffic, or resilience problems—not as a ranking shortcut—and set a cache policy for each content type.
Write alt text around an image's purpose on the page: convey useful product information, name linked actions, explain complex data nearby, and silence decoration.
Most affiliate sites should serve complete article HTML first and use JavaScript for enhancement. Reserve client-side rendering for interactions that truly need it.
Use a 301 only when an affiliate URL is permanently replaced by a useful on-site page. Prefer server-side redirects, keep the map one-to-one, and avoid homepage dumps.
Affiliate tracking cookies are usually not strictly necessary. Get a real choice before setting them, and do not treat Consent Mode as the consent itself.
Use a circuit breaker to stop calls to an unhealthy AI dependency, fail fast during recovery, and admit only a few probe requests before restoring traffic.
Use a dead-letter queue to isolate AI jobs that exhaust bounded retries, preserve diagnostic context, and replay them only through a controlled, idempotent path.
Do not noindex taxonomy archives by default. Keep useful category hubs searchable, and exclude thin, overlapping, or machine-made tag pages with a clear URL policy.
Last-click attribution gives a conversion to the final eligible tracked referral before the sale—not necessarily the final influence or the last click you can see.
Set one end-to-end deadline for an AI workflow, divide it across steps, cancel expired work, and reconcile uncertain side effects before retrying.
Most affiliate sites should use breadcrumbs when they have meaningful categories or multiple levels, but the trail must reflect real hierarchy and remain accessible.
Lazy-load below-the-fold affiliate images, but keep likely above-the-fold visuals eager and preserve crawlability, layout stability, and image quality.
Affiliate links can still track during a private session, but temporary storage, cookie blocking, browser changes, and program rules make credit less durable.
Infinite scroll can work on an affiliate site only when crawlable pagination, stable URLs, accessible controls, and reliable back-button behavior exist underneath it.
Use hreflang when an affiliate site has equivalent, indexable pages for different languages or regions—not merely different offers, currencies, or tracking links.
Protect PII in AI workflows by minimizing inputs, isolating identity, controlling every processor, limiting retention, and testing each boundary for leakage.
Most affiliate sites should use an XML sitemap, but it is a discovery and monitoring aid—not a shortcut to indexing or rankings.
Affiliate coupon codes can apply a discount, identify a referral, or do both—but the program's credit rules decide whether a valid sale earns commission.
Google no longer shows FAQ rich results, so most affiliate sites should not add FAQPage markup for Search alone; keep useful answers and audit old code.
An affiliate program sets a merchant's offer and rules; a network can provide the marketplace, tracking, reporting, and payment rails around many programs.
Affiliate comparison tables work when they compress verified decision factors—not when they replace evidence, bury disclosures, or collapse on mobile.
Treat API rate limits as a scheduling constraint: control admission, share capacity fairly, retry safely, and preserve work when capacity runs out.
Most affiliate sites should keep robots.txt simple: let important articles be crawled, list the sitemap, and block only verified low-value URL patterns.
Affiliate links identify a referral and route a reader to a merchant, but program rules—not the URL alone—decide whether a later action receives credit.
Treat every model output as a proposed action: validate its structure, meaning, authority, and destination handling before anything changes.
Author bios help readers understand who made an affiliate recommendation, but they work only when the identity is accurate and the page carries real supporting evidence.
Treat prompts as deployable behavior: pin immutable revisions, test the whole behavior bundle, stage releases, log the resolved version, and keep rollback boring.
Keep AI workflow API keys out of prompts, code, clients, and logs. Store them centrally, retrieve them at execution time, scope access, rotate safely, and rehearse revocation.
AI workflows need an idempotency boundary around every side effect. Use stable action keys, atomic claims, replayed results, bounded retries, and reconciliation to stop duplicate work.
Affiliate deep links send readers to a specific product, category, or content page through an approved tracking link. Build them cleanly, test both routing and attribution, and maintain the destination.
EPC divides commission by clicks, but the number is only comparable when its time window, click definition, validation status, currency, and traffic source match.
An out-of-stock offer does not automatically justify deleting or redirecting the page. Verify the status, protect the reader, then choose the right URL action.
Most affiliate sites should publish a privacy notice once they use analytics, email, ads, or tracking. Start with the site's real data flows, not a generic template.
Production monitoring for AI workflows needs traces, task-level evaluations, risk signals, user feedback, and rehearsed responses—not one generic accuracy score.
Test the complete AI workflow—not just its final prose—with representative cases, step-level assertions, sandboxed tools, failure drills, and explicit release thresholds.
Prompt injection cannot be solved by a stronger prompt alone. Reduce exposure with trust boundaries, limited tools, approvals, validation, testing, and logs.
There is no universal safe number of affiliate links per page. Keep links that help a reader act, compare, verify, or choose—and remove redundant pressure.
Server-to-server tracking sends conversion data from an advertiser’s backend to an affiliate network, but it still needs a durable click identifier and careful validation.
Affiliate sub IDs are publisher-defined labels that connect clicks and conversions to a channel, page, placement, or creative without changing the offer itself.
YouTube permits affiliate content, but safe use depends on clear in-video and description disclosures, useful original content, approved destinations, and program terms.
Most affiliate pages should declare one clean, self-referencing canonical URL. Use canonicals for true duplicates—not to erase useful reviews, comparisons, or filtered pages.
Affiliate links do not automatically damage rankings. The real risks are unqualified paid links, thin content, weak evidence, and pages built around clicks instead of decisions.
A practical system for routing AI work by risk, giving reviewers useful evidence and real authority, and learning from overrides instead of rubber-stamping output.
A practical audit for finding failed, redirected, mistargeted, and commercially dead affiliate links—then fixing the ones that hurt readers first.
A practical checklist for using affiliate links on Pinterest without hiding the commercial relationship, surprising readers, or drifting into repetitive spam.
A practical rule for choosing same-tab or new-tab affiliate links without confusing readers, weakening accessibility, or mixing up disclosure and SEO requirements.
A practical guide to using Product and Review structured data on affiliate pages without inventing ratings, mislabeling roundups, or chasing guaranteed rich results.
A practical check for program terms, Google Ads destination rules, trademark limits, landing-page value, disclosures, and tracking before buying traffic.
A practical way to check program terms, email-platform rules, recipient permission, disclosures, and link behavior before sending affiliate offers.
A risk-based refresh system for affiliate reviews, comparisons, and guides—built around evidence changes, offer volatility, policy shifts, and reader impact.
A practical guide to deciding whether a branded affiliate redirect is worth using, with program-policy, disclosure, security, tracking, and maintenance checks.
A practical guide to pending, approved, locked, and reversed affiliate transactions—and the evidence to check before accepting or disputing a change.
A practical guide to reading affiliate tracking windows, overwrite rules, browser limits, cart exceptions, and the reports that actually determine credit.
A practical GA4 setup for separating all outbound clicks from affiliate clicks, preserving useful parameters, testing cleanly, and respecting attribution limits.
A practical guide to tagging affiliate links for Google, separating machine-readable rel attributes from reader-facing disclosures, and auditing the rendered page.
A practical, human-in-the-loop system for automating TikTok research, production, approvals, and posting without crossing into spam or fake engagement.
A practical method for turning real testing, sourced facts, clear disclosures, and honest tradeoffs into an affiliate review that helps a reader decide.
A plain-language system for deciding when affiliate disclosures are needed, where to place them, what to say, and how to check each format before publishing.
Before comparing commission rates, run every affiliate program through five pass/fail gates—audience fit, readable terms, allowed channels, clear attribution, and workable disclosure. Fail a gate, move on.
A practical, no-hype sequence for starting affiliate marketing by building one useful income asset, disclosing clearly, measuring behavior, and improving it.
Most affiliate sites chase rankings. This system starts with reader decisions, maps intent clusters to page roles, and uses Search Console signals to drive disciplined refreshes.
AI can accelerate bounded affiliate content tasks, but a human must remain responsible for sources, claims, evidence, disclosure, fact-checking, and the final publish decision.