Technique Graph

1. Remove the 301 redirects that point from the old site to the new site.
2. Add 301 redirects from the new site back to the old site.
3. Open the Change of Address tool in the old property, and select Cancel Move.
4. Plan the reversal with care: a second move cannot be submitted right after the first.

Search Console Help allows the cancel for 180 days, and warns that Google might continue to redirect if the old directives stay live. The reversal makes Google recrawl both sites.

1. Measure the HTML size of each template that must rank.
2. Measure every CSS and JavaScript bundle too: each file is fetched separately under the same limit.
3. Move bulky inline data out of the HTML, and split any bundle that is near the limit.

Google's Googlebot documentation says it crawls the first 2 MB of a supported file type, and the first 64 MB of a PDF. The 2 MB figure replaced the older 15 MB, so pages that were safe before can now be cut off.

1. Split long articles into sections of about 375 words or fewer.
2. Give each section a question as its H2.
3. Keep a clean heading hierarchy, because parent headings can travel with a chunk.

Google's Discovery Engine, the retrieval layer of Vertex AI Search, allows chunks of up to 500 tokens, about 375 words. Google's consumer AI features likely chunk in a similar way, so a section under the cap can fit in one chunk.

1. Open AI Performance in Bing Webmaster Tools, then Grounding Queries.
2. List the phrases Copilot generated when it retrieved your pages.
3. Add a section for each sub-question the page does not answer.
4. Read the citation trend by quarter, not by month.

This is first-party data on the queries an AI engine runs, and Google gives nothing like it. Monthly counts swing too much to read: one property fell from 1,520 citations a day to 34 over 2 months with no known content change.

1. Send a prompt about something current, so that ChatGPT searches.
2. Open DevTools, then Network, and filter by the conversation id from the URL.
3. Reload, open the conversation response, and search it for 'queries'.
4. If the payload changes, use Perplexity: it prints its searches on the page.

ChatGPT usually runs 1 to 3 searches per prompt, and their wording is not the prompt's wording. Those search strings are the keywords to rank for. The payload is not documented, so the Perplexity check is the stable fallback.

1. Ask ChatGPT, Perplexity and Google your buyers' deal-breaker questions.
2. Trace each wrong claim to its source, often a Reddit thread or an old article.
3. Publish one detailed page with the correct facts.
4. Then corroborate the facts on third-party pages.

In an Ahrefs test, most AI platforms repeated a detailed false story over the brand's own vague FAQ. Detail wins. The feedback buttons on each platform are slow and promise nothing.

1. List every profile that repeats the entity description, such as Crunchbase, LinkedIn and review sites.
2. Change all of them within the same few days.
3. Watch the Knowledge Panel subtitle for the update.

The one documented before-and-after case: Jason Barnard changed 61 profiles in one afternoon in January 2022, and his Knowledge Panel subtitle updated in 3 weeks. The same kind of change took 4 to 6 months when it was spread out.

1. Call the kgsearch entities:search endpoint with the brand name and types=Organization.
2. Record whether a result carries an @id.
3. Query each competitor in the same run.
4. Divide each resultScore by the highest score in that run, and track the index by month.

A missing @id suggests that Google has not resolved the brand as an entity. Raw scores vary by query, so only the index inside one run compares across months.

1. Open google.com/preferences/source?q= followed by your bare domain to confirm the site is eligible.
2. Use that URL as the deeplink on a button in the header, article ends and newsletter footer.
3. Fire a GA4 event on each click.

A reader's preferred sources appear more often in Top Stories, so this pays for sites that publish fresh, news-type content. No Search Console or GA4 metric counts your selections, so the click event is the only measure.

1. List the 5 tokens: meta-externalagent, meta-externalfetcher, meta-webindexer, meta-externalads and facebookexternalhit.
2. Decide per token, because they cover training, live fetch, index, ads and link preview.
3. Block meta-externalfetcher at the WAF if you need it blocked: it is user-initiated and may bypass robots.txt.
4. Allow 24 hours for a robots.txt change, because the crawlers cache the file.

Most guides list only 2 of the 5. A block on the common pair leaves meta-webindexer open, which builds the index behind Meta AI answers. Meta's own crawler documentation describes all 5 and the fetcher's robots.txt exception.

1. Read the Signature-Agent header, and fetch /.well-known/http-message-signatures-directory at that host.
2. Pick the key that matches keyid in Signature-Input, and verify the Signature.
3. Check the created and expires values to reject replayed requests.

Web Bot Auth is an IETF draft built on RFC 9421 HTTP Message Signatures, so a signed request proves identity without an IP list. Cloudflare Browser Rendering signs its requests today. Most AI crawlers still identify only by user agent and IP range, so keep the IP check for the rest.

1. Check whether Bot Fight Mode is on for the zone.
2. Write the allow rule to match the user agent and Perplexity's published IP list together, and place a Skip rule above your blocking rules.
3. If Perplexity is still challenged, decide whether to turn Bot Fight Mode off. That removes its protection for the whole zone.

Perplexity's bot documentation says 'Cloudflare's Bot Fight Mode can't be skipped with WAF custom rules', so the mode can still challenge its bots after a correct allow rule.

1. Verify the section, such as a subfolder, as its own Search Console property.
2. Open Settings, then Search generative AI, and set that property to Exclude.
3. Leave the parent property on Include, and allow 1 to 2 days.

Search Console Help says a property inherits from its closest parent that changed the control, so a child property can hold its own setting. Exclude covers AI Overviews, AI Mode and Discover AI features together, and Google says an excluded site gets no traffic or impressions from them. It does not affect AI training or ranking elsewhere.

1. List the pages that carry max-snippet.
2. Audit the text those pages supply in structured data, because Google treats it as separately permitted.
3. Remember that max-snippet:0 equals nosnippet, and that a robots.txt disallow means the tag is never read.

Google's robots meta tag documentation says max-snippet also limits how much content may be used as direct input for AI Overviews and AI Mode. It adds that the limit does not apply where a publisher has separately granted permission for use of the content.

1. Open Admin, then Attribution settings, and note the key event lookback window. The default is 90 days; 30 and 60 are options.
2. Before you shorten it, export session source and medium for returning users.
3. Compare the same report after the change.

Google Analytics Help says the window you choose 'also applies to session attribution'. Google's example is a user who returns on day 68. Inside a 90-day window that session keeps its campaign source, so a shorter window should send it to Direct.

1. Open the key events list after the link.
2. Look for 7 events marked automatically: add_to_cart, add_to_wishlist, begin_checkout, session_start, view_item, view_item_list and view_search_results.
3. Check that none of them is imported into Google Ads as a conversion you bid on.

Google Analytics Help says that with Ads Personalization on and a linked Google Ads account, 'these events are also marked as key events'. Key event totals rise with no change in user behavior.

1. Find Google Ads URLs where a template adds only 1 or 2 UTM parameters.
2. Either remove the UTMs and rely on auto-tagging, or add utm_source, utm_medium and utm_campaign together.
3. Check paid rows for (not set) after launch.

Google Analytics Help says that when the GCLID cannot be used and any UTM is present, it derives all cross-channel source values 'from UTM parameters, exclusively'. Each missing parameter then reports as (not set).

1. Divide Yahoo feedback loop complaints by mail delivered to Yahoo domains.
2. Treat the result as a floor: Yahoo divides by mail that reached the inbox, which you cannot count.
3. Keep that floor well below 0.3%.

Yahoo Sender Hub says 'Spam rate is calculated in our system based on mail delivered to the inbox'. Mail filed to spam leaves Yahoo's denominator, so Yahoo's rate is higher than a rate based on all delivered mail.

1. Publish p=quarantine with t=y first.
2. Read the reports: with t=y, receivers apply the policy one level below the one you published.
3. Remove t=y when the reports are clean, then repeat the steps for p=reject.

RFC 9989 removes the pct tag and defines t=y as a test mode, where the applied policy is 'one level below the specified policy'. A receiver that still follows the older RFC 7489 ignores unknown tags and can apply the full policy, so the safe start is quarantine, not reject.

1. Find the mechanisms in your SPF record that return no DNS answer.
2. Remove or fix each one, so that void lookups stay at 2 or fewer.

RFC 7208 says implementations should limit void lookups to 2, and a check that goes over the limit returns permerror. This limit is separate from the limit of 10 DNS lookups, so a record can pass that count and still fail.

1. List every third party that received the consumer's personal information by sale or sharing.
2. Notify each one of the opt-out, and direct it to comply and to forward the request.
3. Stop selling or sharing as soon as feasibly possible, and within 15 business days at most.

11 CCR § 7026(f)(2) requires this of a business under the CCPA that sells or shares personal information. Unlike the deletion rule, it has no exemption for impossible or disproportionate effort. The duty runs to third parties, not to service providers or contractors under a CCPA contract.

1. Treat a Global Privacy Control signal as an opt-out of sale and sharing for that browser or device.
2. Apply it to every profile linked to that browser or device, including pseudonymous profiles in a CDP or identity graph.
3. Show the visitor that the signal was processed, for example with the text 'Opt-Out Request Honored'.

The CCPA regulations (11 CCR § 7025(c)(1)) extend the signal to any consumer profile associated with the browser or device, including pseudonymous profiles. A posted Do Not Sell link does not replace processing the signal (§ 7025(e)). The display of the processed status applies from 2026-01-01.

1. Check whether any payment changes the order of a comparison or roundup page.
2. Remove any payment that sets the order.
3. Do not rely on a disclosure to cover a paid rank.
4. In the EU, disclose any paid placement on the results surface itself.

In the US, the FTC Endorsement Guides (16 CFR § 255.4(b), Example 3) call paid-for rankings deceptive whatever the site says about independence, and say a disclosure is inadequate when payments decide the ranking. The manufacturer who pays can also be held liable. The EU rule is point 11a of Annex I of the Unfair Commercial Practices Directive.

1. Calculate the average star rating with and without the incentivized reviews.
2. If they raise it materially, place a disclosure next to the displayed average.
3. Tell reviewers in the request that a negative review is acceptable.

In the US, the FTC Endorsement Guides (16 CFR § 255.5(b), Example 6(ii)) say the practice could still be deceptive, even with a disclosure on each review, when incentivized reviews are counted in an average star rating and raise it materially. The example gives no threshold for 'materially', and it assumes reviewers know that a negative review is acceptable.

1. Tie the data study to one product category.
2. Link the category page in your own press release, where the credit goes.
3. Credit the brand and the category together, such as OnBuy Pet Suppliers, and let the journalist choose the final wording.
4. Measure the share of links that hit money pages before and after.

Journalists usually copy the link pattern of the release. Rise at Seven's work for Dojo moved 54 links from one exclusive story into the card-machine category page, and that category's traffic rose 152%. Before the program, 16.5% of Dojo's links pointed at category pages.

1. Search whatdotheyknow.com first: an earlier successful request for the same data saves the wait.
2. Number each question, and structure it as what, where, when, why and how.
3. Write dates as DD/MM/YY, and ask for a spreadsheet by email.
4. Leave out every reference to Acts or laws.

A UK public body has 20 working days to answer, and a clarification question resets that timer. On a campaign with a 4 to 8 week cycle, one vague question can use up the whole buffer.

1. Split the prospect list into campaigns of 51 to 100 emails.
2. Write each campaign's angle for that group.
3. Scale with more campaigns, not larger sends.

In BuzzStream's analysis of 31 million outreach emails, campaigns of 51 to 100 emails earned the most links, 12.08 on average. Campaigns of 1,001 to 5,000 emails earned about 70% fewer. The data is correlational.

1. Export backlinks with the title and surrounding-text columns.
2. Group rows with the same article title and text, and count each group once.
3. Map domains that belong to one publishing group, such as Reach PLC titles.
4. Report 4 numbers: backlinks, referring domains, deduped domains and syndicated copies.

One article syndicated across a newspaper group is one editorial decision. As a worked example, 60 reported referring domains at a 33% syndication rate is about 40 real placements. Ask vendors to quote on the deduped number, with syndication stated.

1. For follow links, start with Source of Sources: 36.26% of its requests offer one.
2. For strong sites, use Qwoted: 70.3% of its requests come from DR 70 to 100 sites.
3. Screen HARO requests: just over 38% come from DR 0 to 9 sites.
4. Check each request for a follow link before you pitch.

BuzzStream's study of 7 journalist request platforms found that only 21.9% of requests explicitly offer a follow link, and that the platforms differ widely. The data is from July 2025.

1. As the host site, ask: do we set the editorial agenda for this content?
2. Ask: does the content fit our own first-party topics?
3. Ask: would the placement make sense with no ranking value?
4. Decline when any answer is no.

The test targets content placed on a host only to use its ranking signals, which is what Google's site reputation abuse policy covers. Google has enforced that policy with manual actions on the host's affected sections.

1. Remove or noindex the pages named in the site reputation abuse manual action.
2. Do not move the same content to another folder on the same domain.
3. State the number of removed pages in the reconsideration request.

Google's guidance on the policy says that moving the content within the same domain can be treated as an attempt to get around the policy, which can lead to a wider action.

1. Create a custom channel group with an AI channel.
2. Add 2 conditions joined by OR: medium exactly matches ai-assistant, or source fully matches a regex of escaped hostnames such as chatgpt\.com.
3. Leave x.com and t.co out; claim only x.ai, grok.com and grok.x.ai for Grok.
4. Drag the channel above Referral and Organic Search.

GA4 assigns the first channel that matches, so a channel below Referral gets nothing. A loose pattern catches mail.google.com, and x.com in the list empties Organic Social. The medium condition catches the hosts Google recognizes that your list does not name.

1. Log the path, user agent and referer of every request.
2. Count the landing-page document requests whose referer host is on your AI hostname list.
3. Subtract the sessions GA4 puts in the AI channel for the same period.
4. Report the gap as your own undercount.

Published estimates of AI visits that GA4 cannot see run from 35% to 70%, and they come from vendors. The gap between your logs and your GA4 channel is your own number. Visits with no referrer at all, such as those from free ChatGPT users, are missing from both counts.

1. Count requests per URL per day from ChatGPT-User, Claude-User and Perplexity-User.
2. Chart the daily count for each money page.
3. Investigate a spike: that page is being read into live answers.

These agents fetch a page at the moment a person's question needs it. The count is the closest free, real-time signal of citation, and it needs no tracking vendor.

1. Log every citation of your domain for your money prompts.
2. Mark it named when the answer text names the brand, and ghost when it does not.
3. Report the 2 counts separately, and judge programs on the named count.

In a Semrush study with Kevin Indig, about 6 in 10 AI citations never named the brand: the page was a source, and the answer did not mention it. Only about 13% of appearances were both cited and named.

1. Ask the vendor for the denominator, the sample size and the fieldwork date.
2. Label top-source-pool shares and all-citation shares in separate columns.
3. Use the all-citation share for budget decisions.

Ahrefs gives YouTube 21.1% of Google AI Overview citations in its top-50 pool, and 5.6% of all citations. Reddit shows a similar gap in other studies. The 2 figures lead to different budgets.

1. Ask your tracker whether it samples the API or the web interface.
2. Measure citations on the web interface or on Perplexity.
3. Use the API only for mention rate, and never put both on one chart.

In a SurferSEO study of 1,000 prompts, ChatGPT API answers and web answers shared 24% of brands and 4% of sources. The API does not show what buyers see.

1. In non-English markets, flag every cited Reddit URL that carries the ?tl= parameter: it is a machine-translated page.
2. Report original and translated citations as 2 counts, per engine and per market. Do not drop either.
3. Check that your tracker keeps URL parameters, or the split is lost.

Peec analyzed 64.77 million Reddit citations across 20 countries from March to June 2026. On Google AI Overviews and AI Mode, translated pages were 40% to 73% of Reddit citations in non-English markets, and under 1% in English-speaking markets. ChatGPT's share fell from 6.14% in April to 0.30% in early June, and Perplexity cited none, so the answer differs by engine.

1. Count prompt runs as n, not prompts.
2. Calculate 1.96 times the square root of p(1-p)/n.
3. Treat a change smaller than that margin as noise.
4. Print n next to every rate.

At a 20% mention rate, the margin is 11.1 points at 50 runs and 3.5 points at 500 runs. Repeated runs of one prompt are correlated, so the true margin is wider: treat this number as a minimum.

1. Plan about 30 to 50 queries on Gemini for an interval of plus or minus 5 points.
2. Plan about 90 to 100 queries on Perplexity for the same interval.
3. Plan at least 150 on ChatGPT search, where results converge unevenly.
4. Report a 95% bootstrap interval over responses, with 1,000 resamples.

The numbers come from the arXiv paper 'Quantifying Uncertainty in AI Visibility' (Sielinski, 2026, arXiv:2603.08924). Citation share is a ratio, so the paper rejects the usual Wilson interval for proportions and uses a bootstrap.

1. Check whether view-through conversion optimization is on for each Demand Gen campaign. It is off by default.
2. If it is on, compare the Conversions column with the lead count in your CRM.
3. Expect a gap: Google does not export these impression-based conversions to your backend, and offline conversion import does not cover them.

Google Ads Help lists Demand Gen as an exception where view-through conversions enter the Conversions column once that optimization is enabled. A view counts when at least 1 pixel of the ad is on screen for any time. The feature is an open beta on YouTube and Discover, so its scope can change.

1. In Podscribe, switch the report between modeled and unmodeled results.
2. Report observed and modeled conversions as 2 separate numbers.
3. Check the total against promo codes or vanity URLs.

Podcast pixel attribution matches only listeners on residential IP addresses, and the vendor scales the count up for the rest. In Podscribe's own worked example, 20 attributed purchases at a 47% residential share become 42.6, so more than half of the reported total is inferred. Podscribe publishes no sample size for its modeling basis.

1. When you hash emails yourself for enhanced conversions (Google Ads API or Data Manager uploads), trim whitespace and lowercase every address.
2. Remove dots and plus suffixes for gmail.com and googlemail.com only.
3. Keep dots and plus suffixes on every other domain.
4. Do not hash the country, state, city or zip values.

Google's Enhanced Conversions for Leads guidance sets these rules, and a wrong hash causes a missed match. The failure is silent, so it shows only as a lower match rate.

1. Check the event_time of every server event before you send a batch.
2. Remove any event older than 7 days, and log the events you drop.
3. Send the rest.

Meta's Conversions API documentation says that one event_time older than 7 days returns an error for the whole request, and none of its events are processed. One stale row can lose a full batch of good conversions.

1. Upload each conversion with its GCLID, conversion name and conversion time.
2. Give each separate conversion on the same GCLID its own conversion time.
3. Correct a value by adjustment within 55 days. A retracted conversion cannot be adjusted.

Google Ads deduplicates uploads on the combination of GCLID, conversion name and conversion time. Two real purchases with the same timestamp become one.

1. Upload conversions matched by user-provided data only for clicks up to 63 days old.
2. Upload conversions with a GCLID for clicks up to 90 days old.
3. Capture the GCLID on every lead, so that a sale closed between day 63 and day 90 can still be imported.

Google Ads Help gives 2 limits: 90 days after the click with a GCLID, and 63 days for enhanced conversions for leads. A long sales cycle that relies on hashed emails alone loses the deals that close after day 63.

1. Check the campaign's Final URL before you add URL exclusions.
2. If the homepage must not serve, choose a more specific Final URL, then exclude the homepage.
3. With a Merchant Center feed, expect that the homepage may still serve.

Google Ads Help says an excluded Final URL 'will still be served and receive traffic for your campaign'. The exclusion only limits expansion to other pages.

1. If pinned Responsive search ad assets must always show, turn off Final URL expansion and URL inclusions.
2. Pin assets only in campaigns where both stay off.

Google Ads Help says pinned Responsive search ad assets 'aren't used if a more relevant URL is chosen'. A legal or brand line that depends on a pin can therefore drop out of the ad.