Your organic traffic drops during a Google update, while AI answers mention competitors and sometimes describe your brand incorrectly. The tempting response is to rewrite everything. That usually destroys the baseline you need to work out what actually changed.
You need a diagnosis before you need a recovery campaign. The practical approach is to separate short-term ranking volatility from page-level relevance problems, entity confusion, and the slower process of becoming a dependable source for AI systems.
Treat an update rollout as an observation window, not a verdict
Core updates change broad ranking systems rather than applying a simple penalty to one page. The December 2025 release was Google’s third core update of that year, and its rollout could take up to three weeks. March and June core updates and an August spam update had already made repeated change an operating condition, not an exceptional event.
If rankings move while a rollout is still active, you don’t yet have a settled result. That doesn’t mean you should ignore the data. It means you should preserve it and avoid attributing every movement to a content defect.
- Mark the timeline. Record the announced start of the update, the pages that changed, and the first date each change became visible. Keep unrelated site releases, migrations, and content edits on the same timeline.
- Rule out faults that cannot wait. Check whether affected URLs still load, remain indexable, return the intended status, and are accessible to crawlers. An accidental noindex directive, broken canonical, blocked resource, or server failure should be fixed immediately.
- Segment the movement. Break the loss down by page type, topic, query intent, country, device, and branded versus non-branded demand. A sitewide average can hide one damaged template or one declining topic cluster.
- Save the pre-edit baseline. Export page and query data before changing titles, copy, internal links, or templates. Without that record, you cannot distinguish recovery from normal volatility.
- Delay broad conclusions until the rollout settles. Continue publishing and fixing verified defects, but postpone mass rewrites, deletions, and structural changes made solely in reaction to daily ranking movement.
Read the metrics as clues, not diagnoses. Falling impressions and positions across a related group of pages point toward a relevance or competitiveness problem. Stable positions with fewer clicks call for a closer look at result presentation, query demand, and search features. One template disappearing while the rest of the site holds steady calls for a technical check before a content review.
Google’s standing position is that a core-update decline does not automatically mean a page is defective and that there is no single recovery action. Improvements can be recognized between core updates, although larger changes may become visible after a later update. Set expectations accordingly: make changes because the diagnosis supports them, not because an update created pressure to look busy.
Diagnose search, entity, and AI visibility separately
Search visibility now depends on three connected systems that operate at different speeds. Traditional search engines retrieve current web information. Knowledge graphs organize facts about entities and their relationships. Large language models synthesize information into conversational answers. A brand can be healthy in one layer and weak in another.
The operating horizons are different as well: search improvements may affect near-term discovery, knowledge-graph education can take months, and durable representation in LLM knowledge can take years. Treating all three as one SEO score produces bad priorities.
| Visibility layer | Question to answer | Evidence to inspect | Best next move |
|---|---|---|---|
| Traditional search | Can the right page be crawled, understood, and ranked for the current query? | Indexing, impressions, positions, clicks, affected queries, page groups, and competing results | Repair technical access, intent alignment, content usefulness, or internal discovery |
| Entity and knowledge graph | Can systems identify the organization, people, products, and relationships correctly? | Conflicting names, descriptions, ownership details, profile facts, structured data, and third-party corroboration | Establish one canonical fact set and make every machine-readable claim agree with visible content |
| LLM and AI answers | Can an assistant accurately include, explain, cite, or recommend the brand for the relevant task? | Repeatable prompt tests, factual accuracy, brand inclusion, cited pages, and consistency across answer variants | Strengthen the underlying entity record and publish information that can be extracted and supported |
This separation prevents a common category error. If Google still ranks your pages but an AI assistant misstates your company, rewriting a high-performing page around more keywords is unlikely to solve the identity problem. If your brand facts are consistent but a commercial page loses non-branded rankings, an organization-wide entity project should not replace a page-level relevance audit.
AI answers also need their own measurement discipline. Save the exact prompt, model, date, answer, cited URLs, and whether your brand appeared accurately. One favorable answer is an observation, not a trend. Reuse a fixed set of prompts so that changes in wording do not masquerade as changes in visibility.
Repair relevance without chasing the update
Once a decline remains visible after the rollout and technical checks are clean, work at the level where the evidence concentrates. If one topic cluster lost visibility, audit that cluster. If one page type fell, inspect its template and purpose. A domain-wide rewrite is justified only by domain-wide evidence.
- Define the searcher’s job. Write down what the affected query asks the reader to understand, decide, compare, or complete. Then check whether the page performs that job without forcing the reader through a long preamble.
- Compare the promise with the delivery. The title and search snippet create an expectation. The opening, headings, and main answer must satisfy the same intent. A compelling title cannot rescue a page that answers a neighboring question.
- Locate the information gap. Check whether the page gives a direct answer, explains the mechanism behind it, covers the important limitations, and supplies enough evidence for the reader to verify consequential claims.
- Make accountability visible. Show who created or reviewed the content, why that person or organization is qualified, when meaningful changes were made, and where factual claims come from. Treat authority, notability, and transparency as audit questions, not as invented ranking factors.
- Resolve internal competition. When several pages perform the same job, decide which one should be canonical. Differentiate pages that serve distinct intents. Consolidate genuine duplicates carefully, and redirect a retired URL to the appropriate surviving resource rather than simply deleting accumulated value.
- Reduce extraction friction. Use descriptive headings, explicit names, concise definitions, coherent internal links, and structured data that matches what a person can see. Machines should not have to infer whether two slightly different names refer to the same entity.
- Update substance, not timestamps. Correct outdated facts, improve weak explanations, and remove unsupported claims. Changing a date without materially improving the page gives readers and machines no new reason to trust it.
People-first content is not a license to ignore retrieval. A useful page still needs to be accessible, clearly scoped, internally connected, and written in language that makes its main claims easy to identify. Technical clarity and human usefulness reinforce each other.
Avoid using word count as a repair target. More text can make the answer harder to retrieve and harder to trust. Add material only when it closes a real information gap: a missing condition, an unexplained decision, an absent method, or evidence the reader needs before acting.
Build a brand record that AI systems can reuse
Page optimization helps a system retrieve an answer. Entity optimization helps it understand who supplied that answer. You need both. The goal is to create a consistent, corroborated record of the brand rather than repeat a slogan across hundreds of pages.
- Create a canonical fact inventory. Record the preferred organization name, concise description, official domain, principal offerings, relevant people, locations, and important relationships. Mark which page is authoritative for each fact.
- Publish stable identity pages. Your organization, about, author, product, and contact pages should state their purpose plainly. Keep durable facts separate from campaign language that changes frequently.
- Align visible and structured claims. JSON-LD should describe the content on the page, not introduce a second version of reality. Conflicting names, URLs, roles, or descriptions increase ambiguity. Structured data can clarify a trustworthy fact; it cannot manufacture authority for an unsupported one.
- Connect entities deliberately. Make the relationships among the organization, authors, products, services, and subject areas explicit in copy, navigation, internal links, and structured data. Do not rely on proximity or branding alone to communicate the relationship.
- Seek relevant corroboration. Accurate independent mentions, profiles, citations, and references help systems verify that the brand’s self-description is not the only available account. Correct contradictions at their origin when possible instead of adding more duplicate claims to your own site.
- Publish citation-ready knowledge. Give important topics stable URLs, direct definitions, clear methods, named ownership, and inspectable evidence. If a claim is an opinion or company position, label it as such. If it is factual, make the support easy to follow.
- Audit machine representation. Test how search results and AI assistants identify the brand, explain its offerings, and associate it with relevant topics. Log factual errors separately from simple absence: correcting a wrong identity requires different work from earning consideration for a new topic.
This is algorithmic education in practical terms: consistently presenting connected facts that search systems can discover, reconcile, and reuse. It is not a prompt trick, and it does not guarantee inclusion in a model’s training data. Training inclusion is a long-term outcome that you cannot force or confirm from a single AI response.
Your intermediate measures should therefore stay observable. Track whether canonical facts agree across owned pages, whether relevant third parties corroborate them, whether search engines retrieve the intended pages, whether AI answers become more accurate, and whether repeated prompt tests show more stable inclusion. Those indicators won’t prove that a model has learned the brand permanently, but they will reveal whether the evidence environment is improving.
Key takeaways: run one visibility program at three speeds
- During a core-update rollout, preserve your baseline, fix verified technical faults, and avoid broad edits based on unsettled movement.
- Diagnose traditional rankings, entity understanding, and AI-answer visibility as separate layers with different evidence and timelines.
- Apply content repairs to the page type or topic cluster where the loss is concentrated instead of rewriting the whole site.
- Use structured data to clarify visible, supported facts. It is not a substitute for consistent identity, useful content, or outside corroboration.
- Measure AI visibility with a fixed prompt set and a log of models, dates, answers, citations, and factual errors.
- Expect page-level search work to operate faster than knowledge-graph development, while durable LLM representation remains a long-term objective.
Turn this into a routine. During a confirmed rollout, save a daily snapshot without making a daily strategic decision. After the result settles, review affected page groups weekly while improvements are in progress. Check canonical brand facts monthly, and run the same AI prompt set on a regular schedule that your team can maintain.
Start with one important topic cluster. Export its current search baseline, identify whether the failure sits in retrieval, relevance, entity understanding, or AI representation, and make the smallest change that addresses that diagnosis. That gives you a result you can evaluate and a method you can repeat when the next shift arrives.
