When a competitor's article appears as a source in an AI answer and yours does not, “How do we get crawled?” may be the wrong first question. A page can be crawled and never appear. It can be found for a related search and not selected. It can contribute to an answer without receiving the visible citation you expected.
Discovery, retrieval, selection, and citation are different stages. The public evidence available to a law firm is incomplete, but it is still useful when each observation is assigned to the right stage.
That distinction turns a vague visibility complaint into a diagnostic: verify the earliest stage that failed, improve what the firm controls there, and do not claim that evidence at one stage proves the next.
First confirm that the answer used current web search
An AI product may offer several modes and sources. One response may search the public web, another may rely on supplied documents or connected data, and another may answer without showing current web sources.
Record the exact product, mode, date, question, account or location context, answer, and visible sources. If the experience did not use public web search, it cannot test whether the current website was discovered in the same way as a cited search answer.
This sounds basic, but it prevents an expensive category error. “The model does not know our new partner” could reflect a mode that never checked the current web, a public-record conflict, an inaccessible page, weak relevance, or an unexposed selection decision. Those diagnoses need different work.
Stage 1: discovery means the information can be encountered
At the discovery stage, ask whether the relevant system had a viable path to the public information.
Evidence can include:
- a working rendered page;
- indexing and canonical settings that match the firm's decision;
- internal or external links that make the page discoverable;
- a verified crawler request in server logs;
- a platform-specific access setting;
- a search index or webmaster report where the platform provides one.
Crawler names matter. OpenAI documents OAI-SearchBot for search, GPTBot for potential model training, and ChatGPT-User for user-triggered requests. Perplexity documents a search crawler separately from its user-triggered agent. One robots rule or log line does not describe every role. OpenAI crawler documentation; Perplexity crawler documentation.
A crawler request establishes that a resource was requested. It does not show that the content was indexed, retrieved for a particular question, selected, cited, clicked, or responsible for a client.
The practical response to a discovery failure is technical and policy-specific. Confirm the firm's intent, inspect the rendered response, and fix the actual block. Rewriting the article while the intended source is unavailable skips the broken stage.
Stage 2: retrieval means candidate information may match the task
Once information is discoverable, a search process still has to find material relevant to the user's need.
The person's words are not necessarily the only query involved. Google describes query fan-out in its generative Search features: related searches can be generated to gather information for an answer. Google AI-search optimization guidance.
Suppose a founder asks how to choose counsel for acquiring a competitor. Relevant retrieval could involve transaction type, jurisdiction, attorney experience, diligence, deal structure, conflicts, or the difference between deal counsel and general business advice. That list is an explanatory example; it is not a claim that a platform issued those exact searches.
The editorial implication is to build a coherent body of information around the real client decision. It is not to create a separate page for every imagined hidden query. A strong acquisition page can define its scope, explain the process and conditions, connect the relevant attorneys, and point to deeper resources where the reader's task genuinely changes.
Public evidence of retrieval is often limited. A displayed grounding phrase, search result, or cited source can show something about what surfaced. It does not reveal every candidate considered.
Systems evaluate and assemble candidate material before producing an answer. Microsoft has publicly described ranking and selection within its AI-search process. That makes “AI answers do not rank anything” an unhelpful universal claim, even though the interface differs from a conventional list of results. Microsoft AI-search content guidance.
Site owners do not receive a complete explanation of that selection. A page appearing once does not identify the feature that caused it. A competitor's heading, word count, schema, domain, or backlink profile may correlate with an appearance without being the decisive cause.
Treat patterns as hypotheses. A clear comparison table may make a page more understandable to people and systems. That independent value justifies the edit. Claiming the table “forces citations” does not.
At this stage, improve the substance the firm can defend:
- answer the intended question directly;
- keep material conditions beside the claim;
- distinguish verified experience from promotional adjectives;
- identify the appropriate attorney, source, and review date;
- preserve a useful page owner instead of creating near-duplicates;
- connect related information with clear navigation and internal links.
Those changes strengthen the source without pretending to reverse-engineer an undisclosed formula.
Stage 4: citation is a displayed source event
A citation is the visible reference presented with an answer. It may link to a firm page, name an outside source, or appear beside only part of a larger response.
Open it. Compare the adjacent statement with the source.
For a law firm, ask whether the answer identifies the right practice, attorney, office, and jurisdiction; preserves the limits around educational information; and avoids extending a source beyond what it supports.
Citation presence is not proof of accuracy. It is also not necessarily a recommendation. An educational article cited for a definition may have helped answer a question without presenting the firm as counsel.
Record the source and the role it played:
Scroll sideways to review every column.Each row is shown as a labeled card.
| Displayed event | What you can report | What remains unknown |
|---|---|---|
| Firm page visibly cited | The page appeared as a source under recorded conditions | Why it was selected or whether anyone clicked |
| Outside profile cited about the firm | That profile contributed to the observed answer | Whether the website was considered |
| Firm named without a source | A mention occurred | Which information supported it |
| Firm presented as an option | Recommendation-like language occurred | Suitability, referral visit, inquiry, or engagement |
This language may feel conservative. It is also decision-grade. It tells the team whether the next task belongs to content, public-record maintenance, technical review, monitoring, or intake.
Stage 5: post-citation behavior is a separate client journey
The source can be accurate and still produce no visit. A visit can occur and fail to become a qualified inquiry. A prospect can use an AI answer and later arrive through branded search or a referral link.
Website analytics and intake evidence belong after the citation stage, not inside it. Preserve identifiable referral visits, reported research influences, matter qualification, consultations, and retained clients as separate fields.
This prevents the most consequential reporting error: calling a citation a case.
A completed diagnostic for a missing law-firm citation
Consider a hypothetical business-law firm with a detailed acquisition guide. A managing partner sees another firm's article cited in an answer about buying a competitor and asks why the firm's guide is missing.
The team runs this fault tree:
Scroll sideways to review every column.Each row is shown as a labeled card.
| Stage tested | Evidence | Finding | Next action |
|---|---|---|---|
| Experience | Saved answer shows current web citations | The observation is suitable for source analysis | Continue |
| Discovery | Guide returns 200, renders fully, uses the intended canonical, and is reachable internally | No confirmed access failure | Continue; do not promise selection |
| Retrieval relevance | Guide explains the acquisition process but never states which buyer decisions it covers | The source is broad and difficult to match to the founder's task | Add an attorney-reviewed scope and decision section |
| Selection | No platform evidence identifies why the competitor was chosen | Cause remains unknown | Do not copy the competitor or claim a hidden factor |
| Citation | Competitor source supports only the diligence portion of the answer | The opportunity is a narrower diligence explanation, not a wholesale rewrite | Improve the relevant existing section and preserve deal/jurisdiction conditions |
| Client path | Guide links only to a generic contact page | A qualified reader cannot identify the relevant lawyer or consultation context | Add a coherent, approved next path |
The diagnosis does not deliver a guaranteed citation. It prevents three unsupported responses: changing crawler settings that already work, cloning the competitor's page, and rebuilding the entire guide when the observed source played a narrow role.
After release, the team can repeat the same question under comparable conditions and examine platform reporting where available. A later citation would be an observation. A change would not, by itself, prove which edit caused it.
Interpret platform reports narrowly
Bing's AI Performance documentation describes grounding queries, topics, citations, cited pages, and query-specific Citation Share in supported experiences. Grounding queries are not a complete log of every user's wording, and Citation Share does not reveal competitor domains or become a traffic or quality score. Bing AI Performance documentation.
Google's generative AI performance report documents impressions for AI Overviews and AI Mode. Those observations are not retained clients. Google generative AI performance report.
Third-party prompt panels may use a fixed question set, different platforms, and their own scoring formulas. Preserve the sample and method with the number. Two reports can disagree because they observe different experiences rather than because one is wrong.
Use the earliest failed-stage rule
When a source does not appear or appears inaccurately, diagnose in order:
- Was the observation actually a public web-search experience?
- Could the intended system access the right rendered source?
- Does that source answer the relevant client task?
- Is the answer supported, qualified, and attributable?
- What did the displayed citation actually support?
- Is there any post-click or intake evidence?
Act at the earliest failed stage. Fix access before content expansion. Fix question ownership before formatting. Fix unsupported claims before promotion. Improve the client path before buying a larger visibility report.
If every observable stage is sound and the page still is not cited, the honest conclusion may be that selection is unknown. Continue observing only if the question matters enough to justify the effort.
Improve what the evidence can support
AI-search systems do not expose a complete source-selection blueprint to law firms. That uncertainty is a reason for sharper diagnosis, not a reason to do nothing.
A crawler log can confirm access. A rendered-page review can confirm delivery. A source audit can confirm whether the page answers and supports a claim. A saved answer can confirm an observed citation. Analytics and intake can confirm parts of the client path. Each fact supports a useful next move when it keeps its own name.
If your team is dealing with failures across access, content ownership, site relationships, generated-answer evidence, and measurement, Juris Digital's law-firm SEO service is the relevant commercial context. Bring one observed answer, its cited sources, the page you expected to appear, and any access evidence. The first decision should identify the earliest failed stage before anyone prescribes more content or a crawler change.
