Why AI Mentions Drop After a Domain Migration Even When Google Traffic Recovers

A domain migration changes more than where a page lives. It changes the URLs that external systems have stored, the wording they associate with your company, the links that support your identity, and the evidence available when an AI system retrieves an answer. That is why a site can regain much of its Google traffic while appearing less often in ChatGPT, AI overviews, or other answer surfaces.

Search recovery and AI recovery are different events

Google and AI systems do not maintain one shared visibility index. Search engines may use redirects, canonical tags, internal links, backlinks, and ranking history to reconnect a new URL with an old one. A retrieval or answer system may instead depend on a mixture of current crawling, stored documents, third-party pages, knowledge graphs, search results, and model-specific data.

These systems also update on different schedules. A search engine may recrawl a high-value page soon after a migration. A directory, review site, industry publication, or public dataset may continue to describe your company using the former domain for months. An answer system can encounter both versions, treat them as separate evidence, or use a stale source that has not followed the redirect.

The first failure mode: historical URLs no longer resolve cleanly

AI systems and the sources they consult may still hold old URLs for product pages, comparison pages, documentation, case studies, and category guides. If those URLs return a 404, a generic homepage redirect, a redirect chain, or a page with materially different content, the historical association weakens.

A migration can appear healthy in a conventional analytics report while still losing the long tail of pages that supported AI visibility. Search traffic may be concentrated on a smaller number of replacement pages, whereas answer systems may have relied on specific URLs that explained a use case or named a category precisely.

1. Export the complete pre-migration URL list from the old sitemap, analytics, server logs, Search Console, and backlink tools.
2. Request every old URL and record the full redirect chain, final status code, canonical URL, and rendered content.
3. Map each important old page to its closest factual replacement, rather than sending unrelated URLs to the homepage.
4. Check that redirects remain in place and are not conditional on user agent, geography, cookies, or JavaScript.
5. Compare internal links, XML sitemaps, hreflang references, feeds, and structured data for old-domain remnants.

A redirect is a discovery and continuity signal, not a promise that every external system will instantly transfer its stored evidence. Keep high-value redirects active for the foreseeable future, and make the replacement page clearly satisfy the same intent.

The second failure mode: entity signals changed during the move

AI visibility depends partly on whether systems can identify the same organization across documents. A migration can disturb that identity even when the logo and company name look unchanged to a human visitor.

Common changes include a shortened company name, a new product taxonomy, a different description in the Organization schema, changed social profile links, a new headquarters address, inconsistent legal names, or a domain that no longer appears in reputable third-party profiles. Removing an About page or consolidating product pages can also eliminate the language that distinguished your company from similarly named entities.

• Use one consistent preferred organization name, legal name where relevant, domain, logo, social profiles, and contact details.
• Check Organization, WebSite, Product, SoftwareApplication, and Article structured data for conflicting identifiers.
• Preserve meaningful product names and explain renamed products with explicit “formerly known as” language when accurate.
• Keep author, reviewer, publisher, and company relationships clear on editorial and research content.
• Update important third-party profiles, partner pages, marketplace listings, and documentation links.

Structured data helps machines interpret a page, but it cannot compensate for contradictory public evidence. The practical goal is not to add every possible schema type. It is to make the same entity legible across your site and the sources that describe it. The JSON-LD and AI discovery guide covers the technical checks without treating schema as a ranking shortcut.

The third failure mode: redirects preserved URLs but not meaning

Many migrations include a redesign, rebrand, CMS change, or content consolidation. In that process, pages may become shorter, more generic, or more marketing-led. A page that once said exactly who a product serves, what it replaces, and how it compares with alternatives may now say only that it is an “all-in-one platform.”

That wording change matters. Buyer-intent questions often require specific associations: which tools serve a particular team, which vendors support a workflow, or which products fit a constraint. If those terms disappear, the site may remain indexable and rank for broad queries while becoming less useful as evidence in an answer.

Run a meaning-diff, not just a status-code audit

For important old and new URLs, compare the page title, headings, visible HTML, product and category terms, claims, examples, authorship, references, and calls to action. Pay special attention to content that answered comparison or selection questions. Do not restore every old sentence; restore the facts and distinctions that made the page useful.

• List the entities, products, industries, integrations, alternatives, and use cases named on the old page.
• Mark which facts remain, which moved to another URL, and which disappeared.
• Check whether the replacement page is rendered in crawlable HTML rather than requiring a client-side application to load its main content.
• Verify that canonical tags, metadata, sitemap entries, and visible copy describe the same primary topic.
• Test representative pages with the crawlable HTML versus SPA guide.

The fourth failure mode: third-party sources have not caught up

A domain migration does not automatically update the wider web. Review platforms, partner sites, comparison articles, analyst pages, documentation mirrors, job listings, podcasts, community posts, and industry directories may continue to use the old domain. Some sources will follow a redirect; others will preserve the old link or never revisit the page.

This creates a split identity: your new site says one thing, while external sources still provide older names, URLs, descriptions, or product relationships. In some cases, those sources are more likely than your own site to be retrieved for a recommendation answer because they appear independent or directly compare vendors.

1. Prioritize sources that rank for your brand, product, category, and comparison queries.
2. Ask partners and publishers to update the visible domain and destination URL, not merely the link target.
3. Review review-site profiles, software directories, marketplace listings, social bios, and documentation references.
4. Track whether old-domain pages remain indexed or quoted by external pages.
5. Record source changes separately from your own technical fixes; they often move on a slower schedule.

This is also where Common Crawl training presence needs careful interpretation. A crawl snapshot can show that a domain was publicly reachable at a point in time. It cannot establish that a particular model trained on it, retained it, or will cite it in a current response.

The fifth failure mode: access and discoverability degraded quietly

Migration launches often introduce small technical changes with large practical effects: a staging rule copied into production, a new robots policy, blocked assets, a sitemap containing only a subset of URLs, or a JavaScript shell with little server-rendered text. Search traffic can recover through already known URLs while newer crawlers encounter a weaker site.

Check both the domain and important URL classes. A robots.txt file may allow standard search crawlers but block an AI crawler, or a CDN may return different responses to different user agents. Do not assume that a successful browser visit proves that a crawler receives the same content.

• Fetch robots.txt from the canonical domain and inspect user-agent-specific rules.
• Test representative URLs with an HTTP client and a rendered browser, recording status, redirects, headers, and visible text.
• Confirm that XML sitemaps list canonical, indexable URLs and are linked from robots.txt where appropriate.
• Check noindex, X-Robots-Tag, canonical, authentication, rate limiting, and WAF behavior.
• Review robots.txt for AI crawlers and document intentional access decisions.

Access is a prerequisite, not an outcome. Allowing a crawler to fetch a page does not mean the page will be used, cited, or recommended.

A practical post-migration diagnosis sequence

Avoid changing five variables at once. A staged investigation makes it easier to identify whether the issue is URL continuity, content meaning, source coverage, or live-answer behavior.

1. Define the visibility loss. Compare pre- and post-migration prompts by mention rate, recommendation rate, citation rate, and competitor share of voice. Keep prompt wording, location, model, and date controls consistent where possible.
2. Build an old-to-new URL map. Include high-value pages and long-tail pages, then test redirect status, destination relevance, canonicalization, and content continuity.
3. Run an entity consistency check. Compare names, products, identifiers, profiles, structured data, authors, and company descriptions across owned and important external sources.
4. Run an access and readiness scan. Check robots rules, HTML rendering, metadata, JSON-LD, sitemaps, llms.txt if used, and response behavior. The AI visibility measurement guide explains how to keep these measures distinct.
5. Run a content meaning-diff. Identify missing category language, use cases, comparisons, evidence, and product relationships rather than focusing only on word counts.
6. Recheck live prompts after a defined interval. Record which pages and sources appear, not just whether your brand name appears.

What not to conclude from a temporary AI visibility drop

A lower mention rate after migration does not automatically mean a penalty, a failed rebrand, or a need to publish dozens of new pages. AI outputs are variable, and a small prompt set can produce noisy results. Compare enough repeated observations to identify a pattern, and separate changes caused by the migration from normal model or source fluctuations.

Likewise, adding an llms.txt file, placing a brand name in every heading, or adding unsupported schema will not repair missing redirects or contradictory sources. Use the llms.txt guide for a measured implementation, and treat it as navigation and context—not a control panel for model behavior.

The strongest recovery work is usually unglamorous: preserve useful URLs, make replacements equivalent in meaning, keep entity details consistent, restore crawlable content, and update the external pages that answer systems may retrieve.

A repeatable recovery plan for the next 30 days

FAQ