How AI Search Engines Decide Which Firms to Cite

An AI system deciding whether to name a law firm is not running a popularity contest. It is running an accuracy filter, and the filter behaves differently depending on what kind of question was asked. Ask a general legal question, "how does divorce work in Indonesia," and most engines will answer directly, citing a source. Ask the same engine to apply the law to your specific facts, "will I win my case," and it will refuse, redirect to a licensed advocate, and cite nobody at all. Understanding that boundary, and building content that sits deliberately on the citable side of it, is the entire discipline of legal GEO.

The Three-Layer Architecture That Determines Citation Odds

Legal GEO content operates across three layers, each carrying a distinct AI citation mechanism and a distinct compliance exposure profile. The relationship between the two is, counterintuitively, inverse: the layer with the lowest compliance risk is also the layer with the highest citation opportunity.

The Query Taxonomy Behind the Layer Model

Academic research, specifically Nguyen et al.'s 2025 study on legal information-seeking queries, categorises the questions people ask into three functional types that map directly onto the three layers above. Rights and status queries, "apa hak saya jika dipecat tanpa pesangon," establish the user's legal situation and get answered with general explanations citing the relevant statute. Process and procedure queries, "bagaimana cara mengajukan gugatan cerai di pengadilan agama," seek a factual, sequential answer and represent the single highest-citation-opportunity category, because the answer is precisely factual and requires minimal professional judgment. Specific advice queries, "haruskah saya menerima tawaran settlement ini," require applying law to one person's facts, and trigger a strong professional redirect with no source cited as authority for the answer.

The practical implication is directive rather than descriptive. A firm building content for AI citation should deliberately overweight process and procedure content relative to situation-specific content, not because the latter is unimportant to clients, but because it is structurally the content type AI systems are least willing to cite regardless of how well it is written.

Why Engines Do Not All Behave the Same Way

Platform-level differences matter for prioritisation, even though exact percentages vary across the audits that measure them and should be treated as directional rather than precise. Consistently across the research, more conservative engines, notably Claude, favour authoritative general content and are the least willing to name a specific firm for a recommendation-style query, which makes them most responsive to Layer 1 educational content specifically. Less conservative engines produce longer, more willing-to-name answers but also disclaim more heavily on the same query types. Directory-referencing behaviour also varies: some platforms lean heavily on aggregator citations for firm-recommendation queries, which reinforces why directory profile completeness functions as GEO infrastructure rather than a side activity.

The Hallucination Problem Is Also a Content Opportunity

US courts recorded 487 documented instances of AI-generated hallucinations in legal filings in 2025, more than ten times the 2024 total, with licensed attorneys accounting for roughly 38% of the problematic filings. That statistic is usually presented as a warning, and it is one. It is also, read from the content-strategy side, a genuine GEO opportunity: AI systems are increasingly discriminating in favour of content that cites verifiable primary sources, because that citation structure is exactly what distinguishes reliable from unreliable content during inference. An article that cites UU No. 18/2003 Pasal 22 directly is structurally more citable than one making the identical substantive point without the citation, because the citing article gives the AI system something it can independently verify.

The scale of the underlying reliability problem is larger than the court-filing count alone suggests. Stanford RegLab and HAI's "Large Legal Fictions" research, the most frequently cited academic source on this question, found general-purpose large language models hallucinating on legal queries at rates variously reported between roughly 58% and 88% depending on the specific test set and model version, while legal-specific tools such as dedicated case-law research platforms still hallucinate at a documented 17% to 34%, lower but far from negligible. Google's own AI Overviews, according to cross-market research, generate a response for approximately 51% of tested YMYL queries, meaning roughly half of all Your Money or Your Life questions now receive a synthesised answer rather than a list of links. Separately, an estimated 58.5% of legal-related searches now end without any click to a website at all. Read together, those three figures describe a search environment where an AI-generated answer, not a webpage, is now the default outcome for a majority of legal queries, and where that answer carries a materially higher error rate than most readers assume.

This is the same mechanism behind the widely cited Princeton KDD 2024 finding that citation-dense, well-sourced content earns roughly 40% more AI citations than unoptimised content covering the same ground. In a category where hallucination risk is documented and rising, the premium on verifiable sourcing is not a minor ranking factor. It is close to the primary one.

The Zero-Cost Step Almost Every Firm Skips

Before any of the content or entity work above can matter, an AI system has to be able to read the site at all. GPTBot, ClaudeBot, PerplexityBot and Google-Extended each need to be explicitly permitted in robots.txt for their respective platform to crawl and index a site's content for training and retrieval. Many law firm websites, particularly those running on conservative default CMS configurations, block some or all of these crawlers without anyone noticing, either through an explicit disallow directive or a catch-all "User-agent: * Disallow: /" rule inherited from a template. Checking and correcting this takes minutes and is the genuine prerequisite for every other GEO activity that follows it.

Why AI-Referred Clients Convert Differently, and What That Means for Reporting

Client data from firms already tracking AI-sourced traffic shows a pattern worth building measurement around specifically: visitors arriving from an AI platform citation convert at meaningfully higher rates than visitors arriving from standard organic search, roughly double in documented cases. The likely mechanism is qualification rather than magic. An AI system that cites a firm's content in response to a specific legal question has already done part of the filtering a search engine leaves entirely to the user, delivering a visitor who arrived already oriented to the firm's relevance rather than one still comparing several open tabs.

This has a direct reporting implication. A firm measuring only aggregate conversion rate loses the signal entirely, because AI-referred sessions are often small in volume relative to total organic traffic and get averaged away. Segmenting conversion tracking specifically by referral source, isolating sessions that arrive from chatgpt.com, perplexity.ai or Google's AI-generated result surfaces, is necessary to see the effect at all, and is the same segmentation discipline recommended for tracking AI citation's business value more broadly.

Content Freshness Is a Citation Risk, Not Just a Housekeeping Task

Legal information decays at a rate set by statutory and regulatory change, and AI systems do not automatically discount a page for being outdated the way a human reader eventually would. Indonesia's Criminal Code reform, phased in through 2026, and the ongoing Cipta Kerja employment law cluster are both live examples of legislation that can render a previously accurate GEO article wrong within months of publication, while the AI system that indexed it may continue citing the outdated version for a considerable period afterward. A quarterly statutory review, scoped by practice area rather than attempted across the entire content library at once, is the practical discipline that keeps a firm from becoming the citation source for law that no longer applies.

Two Gates, Reviewed Together, Not in Sequence

Legal GEO content passes through two review constraints that have no direct parallel in most other content verticals, and the workflow only becomes sustainable once both are checked in the same pass rather than one after the other. The accuracy gate verifies that every factual claim about process, statute or rights is checkable against a primary source. The compliance gate verifies that every claim about the firm's own services clears KEAI Pasal 8(b), 8(f) and 4(c). A third, GEO-specific pass then checks machine-readability: definition-lead paragraphs, FAQ schema implementation, statutory citation by article number, and entity markup completeness, none of which requires an advocate's involvement because it addresses structure rather than substance.

Running all three against pre-cleared templates, rather than assessing each piece of content from a blank page, is what makes the workflow scale. A four-element pre-clearance framework covers the recurring decisions: a content template fixing section order and mandatory disclaimer text, a statutory citation bank of pre-verified quotes checked quarterly against amendments, modular disclaimer components updated centrally when PERADI guidance changes, and an entity markup template for Attorney and LegalService schema fields, approved once at the firm level and updated only when a credential changes. Production against this framework cuts per-article attorney review from roughly ten days to about two, because the attorney is verifying citation accuracy rather than reassessing the entire compliance posture each time.

Measuring Citation Health, Not Just Citation Count

Conventional GEO measurement prioritises volume: how often a brand gets cited, how much referral traffic follows. For legal content, that volume-first approach is incomplete on its own, because someone acting on inaccurate general legal information faces a materially different kind of harm than someone who saw the wrong software recommendation. An accuracy-first framework adds two layers most GEO reporting skips entirely.

A monthly citation accuracy audit samples how AI-generated responses actually represent a firm's content when they cite it, checking whether the platform is paraphrasing correctly rather than simply confirming the citation exists. A quarterly content staleness review checks GEO content against the underlying law it describes, because AI systems can continue citing a page for months after the statute it describes has changed, particularly relevant given Indonesia's ongoing revisions to the Criminal Code and employment law clusters. A high citation frequency paired with a low accuracy score is the specific failure pattern this measurement layer is built to catch before a client acts on it.

LinkedIn as a Second Citation Surface

One citation surface consistently under-considered in legal GEO planning is LinkedIn, which matters disproportionately because Microsoft owns the platform and partners with OpenAI. A well-maintained partner profile, with a credential-specific headline and regular publication of practice-area analysis, appears in AI-generated answers about specific attorneys more often than the same attorney's own firm bio page, particularly for corporate and cross-border practice areas where in-house counsel research in English. Treating LinkedIn as GEO infrastructure, rather than a static professional listing, extends the entity-signal logic described above onto a platform most firms have not yet touched from a citation-strategy perspective.

---

Frequently Asked Questions

If Layer 3 content has the lowest citation value, should firms stop publishing it entirely?

No. Situation-specific content still serves clients directly and supports conversion once a client has already found the firm. The point is not to eliminate it, but to build the content strategy's citation-seeking effort around Layers 1 and 2, where the actual GEO return sits, rather than expecting a Layer 3 page to generate AI visibility it structurally cannot.

Does a firm need to optimise separately for every AI platform?

Largely no. The entity signals, structured data and citation density that earn citation on one platform tend to transfer to the others, because they all reward verifiable, well-sourced content for the same underlying reason. Platform-specific tactics matter most at the margins, not as a substitute for the entity foundation.

How often should a firm check whether AI crawlers are blocked?

Monthly, as part of a standard technical audit, since a CMS update or plugin change can silently reintroduce a blocking rule. It is a five-minute check with an outsized consequence if missed.

Is a high citation frequency always a good outcome?

Not on its own. A citation that misrepresents the firm's actual content, misstates a legal point, or attributes advice the firm never gave is a liability signal, not a GEO success. Citation frequency needs to be tracked alongside citation accuracy, not instead of it.

Given the hallucination rates cited above, should a firm be cautious about GEO investment at all?

The hallucination risk described above is a property of general-purpose AI systems answering questions without a reliable source to draw on, not a reason to avoid being that reliable source. A firm publishing well-sourced, statute-cited content is contributing to the smaller, more accurate side of that statistic, and every well-verified citation an AI system draws from a firm's own content is one fewer instance where it generates an unsupported answer instead.

The layer model above is the same framework behind our Legal GEO service, and it pairs directly with the entity and schema work covered in Legal SEO. For the full platform-by-platform playbook and the citation-accuracy measurement model, see Cited or Silent: The Definitive GEO, AEO & AI Visibility Playbook.

---