Does AI-Generated Content Hurt Your Law Firm’s Search Rankings?

Earlier this month, Anthropic announced that it would switch on invisible watermarking for the text Claude generates. Models released from August 2 forward will now weave a faint statistical signature into their word choices, imperceptible to any reader but detectable by software holding the right key. Google has been doing the same thing to Gemini output since 2024, with considerably less attention and fanfare.

The timing was regulatory rather than technological. European transparency rules took effect the same day.

What followed was predictable. Within days, a small industry of watermark removal tools emerged, and a round of consultant commentary warned that Google is preparing to demote anything carrying a mark. That last claim tends to reach managing partners eventually, usually secondhand and usually stripped of the facts underneath it.

Does Google penalize AI-generated content?

No. Google does not penalize content because a machine helped write it. Its published position is that ranking systems reward quality regardless of production method. What Google does penalize is content produced at scale primarily to manipulate rankings rather than to help a reader, which is a question of purpose and value, not tooling.

This has been Google’s stated position since February 2023, when it published guidance on AI-generated content confirming that its systems reward original, high-quality content demonstrating expertise, experience, authoritativeness, and trustworthiness (EEAT), however that content is produced.

Google’s Danny Sullivan put it more bluntly on social media that same year, correcting a writer who had claimed AI content would not be well received by search engines: their systems look at the helpfulness of content, not how it was produced. 

That cuts both ways. AI content gets no bonus. It also gets no automatic penalty.

The line Google actually draws is in its spam policies, under scaled content abuse. The violation is generating many pages primarily to manipulate search rankings rather than help users. A firm publishing eighty near-identical practice-area pages with the city name swapped is on the wrong side of that line whether a human or a machine wrote them. A firm publishing twelve genuinely researched local-specific articles a year is not.

What is AI content watermarking, and which models actually use it?

An AI text watermark is an invisible statistical pattern embedded during generation. The model subtly biases which words it selects, so a detector holding the right key can later identify the text as machine-generated. Nothing about the writing looks different to a reader. As of August 2026, only Gemini and Claude’s latest models apply this to text.

Google DeepMind developed the underlying method, SynthID-Text, and published it in Nature in October 2024. It has been running on Gemini output since then.

Anthropic followed this month. In its explanation of how Claude’s text watermarking works, the company confirmed it is using a version of the same SynthID-Text approach, and was explicit about why: compliance with the EU AI Act. Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026 alongside roughly 190 other signatories, and applied watermarking globally at launch because there is no durable way yet to scope it by region.

One correction worth making, because the industry chatter has this wrong. OpenAI has not publicly deployed a text watermark for ChatGPT yet. It has built one but so far has not shipped it. OpenAI does attach provenance credentials to generated images and audio, as TechCrunch noted in its coverage of the Anthropic announcement, but images and text are separate issues, and most providers addressed images first.

So as of this writing, just two models watermark text. And one of them is Google’s.

Will watermarked content get demoted in search results?

There is no evidence for this, and the strongest argument against it comes from Google itself. Google has had SynthID running on Gemini text output since 2024. In roughly two years, it has never used a text watermark as a ranking signal, never referenced watermarks in its content guidance, and never suggested marked text ranks differently.

Think about what that means practically. 

The company that invented production-scale text watermarking, open-sourced the method, and has been applying it to its own model’s output for two years has had every opportunity to wire that signal into Search. But it hasn’t.

Google did ship watermark detection at I/O 2026, and this is where much of the confusion originates. Users can now ask whether an image is AI-generated directly in Search through Lens and AI Mode, reading SynthID and C2PA credentials. That is a media provenance feature aimed at deepfakes and manipulated photos. It is not a scoring system for written article text, and Google has announced nothing suggesting otherwise.

Two further points that rarely make it into the scary version of this story. First is that detection tooling for text barely exists. Anthropic has said a detection API is coming but has not released one. There is no public tool for checking OpenAI-generated text at all. Right now, nobody outside these companies can audit your firm’s content for watermarks even if they wanted to.

Second is that the marks are fragile. A watermark survives copy-paste. It does not survive rewriting. Substantial human editing changes exactly the token choices the watermark depends on, which means heavily edited work might not register as marked in the first place.

None of that is a strategy. It is context for how much weight this deserves in your marketing decisions, which, IMHO, is very little.

Should your firm buy a watermark removal tool?

Absolutely no. These products cannot demonstrate that they work, because the detector required to verify removal has not even been publicly released. Most of them strip file metadata or hidden characters, which is not where a text watermark lives. Removing the actual mark means rewriting the text through a second and usually weaker model. 

A market appeared within days of Anthropic’s announcement. BleepingComputer catalogued an open-source project with more than 4,500 GitHub stars, a cluster of newly registered web tools, and at least one established detection-evasion service, all advertising Claude watermark removal. None of those claims can be checked, since Anthropic has not published the detector that would show whether cleaned text still carries the mark. Several commercial sites report scores measured against ordinary AI detectors instead, which test something else entirely.

The more telling admission comes from one of the developers. 

The author of the most-starred project acknowledged that his tool currently removes metadata only, and his own documentation argues against the whole category: a rewrite swaps a premium model’s word choices for a cheaper model’s, which is a strange thing to pay for after paying for the better model.

For legal content, that degradation is the real danger. For example, a paraphrasing pass may not know that Maryland’s Local Government Tort Claims Act notice period changed from 180 days to one year. It will smooth a verified statutory deadline into something that reads well and could be wrong. Every unsupervised rewrite is another chance to reintroduce the errors verification exists to catch.

The exercise is also unnecessary. Anthropic’s documentation states that heavy editing, paraphrasing, and translation all can cause the mark to fade. A genuine editorial pass already does what these products charge for, and it improves the writing rather than degrading it.

Is content marketing still worth the investment for law firms?

It matters more than it did even a few years ago, and here’s why. AI assistants now carry people through an entire decision, from a first vague question to a specific recommendation, and they build that recommendation partly from sources they relied on earlier in the same conversation. Firms that answer the early questions (and the middle questions) are the firms that get cited at the end.

The old pattern was one search, a page of links, and a few open tabs. The new pattern is a conversation that runs the length of the decision.

Semrush surveyed more than 1,000 US consumers about AI and the buying process. It found 57% using AI to narrow their choices, 33% working through multi-turn exchanges that get refined as answers come back, and half reporting a purchase made after researching with an AI tool. People are not stopping at the top of the funnel. They are running the entire funnel inside an AI chatbot.

Legal services sit squarely in that shift. A 2026 survey from iLawyerMarketing found roughly 42% of consumers using ChatGPT to research attorneys, about half using some AI answer engine, and one in ten relying on AI alone. 

Google still leads at 72%, and that is the detail many firms miss. Strong rankings now feed both surfaces at once, so a single content investment, executed properly, works in both traditional search and the emerging AI search engines.

Now there is an important distinction here between getting cited and getting recommended. 

Semrush analyzed ChatGPT across 283,215 citation observations and found that citation share responded to coverage generally, while brand mentions, meaning the model naming a company directly in its answer, tracked depth within a specific category. 

Broad question coverage earns the citations. Depth in one area earns the recommendation. Both are strong drivers of quality traffic. 

Presently, the window of opportunity is unusually open. Semrush’s study of 50,000 brands found only 15.2% of ChatGPT categories have a clear owner, with 89.3% of AI search demand sitting in categories nobody owns yet. Once a brand does own a category, it held that position in 90.4% of month-over-month comparisons. 

For a defined practice area in a defined local market, that is close to an empty field with a compounding reward for arriving first.

The economics reinforce the case. AI referral traffic is small in volume and converts far above organic. Semrush measured AI search visitors converting at 4.4 times the rate of standard organic search. Ahrefs found that AI referrals accounted for 0.5% of its own traffic while producing 12.1% of signups. Fewer visitors, arriving already informed and considerably closer to hiring.

What actually costs law firms visibility in AI search?

The real risk with AI search is irrelevance. AI LLMs retrieve and cite specific passages, not whole pages, and a passage only gets cited if it answers a real question with information the system cannot find in twenty other places. Generic content fails that test regardless of who wrote it.

The mechanics here are worth understanding, because they changed what “ranking” even means.

When someone enters a complex question into Google’s AI Mode, for example, the system does not run one search. It breaks the question into multiple sub-queries and runs them in parallel, a process the industry calls query fan-out. Search Engine Land’s guide to query fan-out traces the mechanism to Google’s own patent filings, where it appears as query variant generation, and notes that essentially every major AI search product now uses some version of it.

The consequence shows up in the citation data. Ahrefs analyzed 863,000 keywords and four million AI Overview URLs and found that only 38% of citations now come from pages ranking in Google’s top 10, down from 76% a year earlier. Roughly two-thirds of citations go to pages outside the top 10, because fan-out reaches deeper into the index to answer narrower sub-questions.

For a firm that has never cracked page one, that is a big opening. A well-answered specific question can earn a citation without a top ranking behind it.

There is academic support for what makes a passage citable. The Generative Engine Optimization paper from researchers at Princeton, Georgia Tech, and the Allen Institute for AI tested nine content strategies across 10,000 queries. Adding statistics, quoting credible sources, and citing authorities produced visibility gains of up to 40%. Keyword stuffing performed worse than doing nothing.

Statistics and citations. That is the finding. Not word count, not keyword density.

Why do question-based blog posts perform better in AI search?

Because retrieval happens at the passage level. AI systems split pages into chunks, and a chunk that directly answers one specific question outperforms a longer chunk that mentions the topic in passing. A question-shaped heading followed by a complete, self-contained answer produces exactly the unit these systems are looking for.

Google documents this behavior openly. Its guide to Search ranking systems describes passage ranking as an AI system that identifies individual sections of a web page to better understand how relevant that page is to a search, and its 2026 generative AI optimization guide confirms that its retrieval-augmented generation systems pull relevant pages and then review the specific information inside them.

Query behavior has shifted to match. Question-format searches trigger AI summaries far more often than keyword-style searches, and the queries people type into AI assistants run dramatically longer than the handful of words they once typed into Google.

The practical structure follows directly:

Lead with the answer. The opening 40 to 60 words carry disproportionate weight in whether a passage gets extracted. 

Make every subheading a real question. Not “Statute of Limitations.” Instead: “How long do I have to file an injury claim in Maryland?” The first matches an outdated keyword report from 2019. The second matches a sub-query.

Make each section survive alone. No “as discussed above,” no pronouns pointing backward. When a chunk is retrieved, the rest of the page does not travel with it.

Put verifiable numbers close to the claims. This is the Princeton finding applied.

What does a generic AI draft get wrong about local law?

Quite a lot, and the errors cluster in exactly the places that decide cases. Take Maryland as an example. A national template will apply comparative negligence rules that do not exist here, cite a superseded notice deadline, and name a court that changed its name four years ago. Each error is invisible to a reader who does not practice in Maryland and obvious to one who does.

Three examples we verify on every relevant draft.

Contributory negligence. Maryland remains one of a small handful of jurisdictions where a plaintiff found even slightly at fault recovers nothing. The Court of Appeals confronted this directly in Coleman v. Soccer Association of Columbia, 432 Md. 679 (2013), and declined to abandon the doctrine, holding that any change should come from the General Assembly. A draft written from national training data will routinely describe fault as reducing a recovery proportionally. In Maryland, it eliminates it.

The Local Government Tort Claims Act notice deadline. If your client was hurt by a Baltimore County vehicle or on Montgomery County property, notice requirements apply before any suit. That deadline is now one year after the injury under Md. Code, Cts. & Jud. Proc. § 5-304. It used to be 180 days. The General Assembly changed it in House Bill 113 during the 2015 session, which also raised the liability caps to $400,000 per individual claim and $800,000 for claims arising from the same occurrence. Search “Maryland tort claims notice” and you will still find pages citing 180 days.

Court names. Maryland voters approved a constitutional amendment in November 2022, and effective December 14, 2022, the Court of Appeals became the Supreme Court of Maryland while the Court of Special Appeals became the Appellate Court of Maryland. Judges on the high court are now justices. Any draft referring to the Maryland Court of Appeals as the current high court is working from stale information.

For comparison, the general personal injury limitations period is three years from accrual under Md. Code, Cts. & Jud. Proc. § 5-101, a deadline that sits comfortably alongside a notice requirement a claimant may have already missed without knowing it existed.

That last combination is the kind of thing a Maryland claimant urgently needs to understand, and precisely the kind of nuance a generic draft gets wrong or omits.

How does Too Darn Loud Marketing use AI in the content process?

We use AI tools as part of the content research and drafting process, with heavy human involvement at every stage. Every statute, deadline, dollar figure, and statistic is verified against a primary source before drafting begins. Claims we cannot verify do not appear. A human editor reviews every piece before a client ever sees it.

We believe this approach puts our clients in the best possible position to succeed. Firms that refuse to use AI tools at all are going to fall behind on volume and turnaround, and firms that use them without a verification layer are going to publish something embarrassing about the law, which could not only damage their reputation, but also land them in trouble with the ABA and local bar associations. Neither is a good place to be.

Our process separates research from writing. The research pass produces a locked set of verified facts with primary-source citations attached. The drafting pass can only use what survived verification. For example, if a claim about Maryland law cannot be confirmed against the Maryland General Assembly, the Maryland Judiciary, or a comparable authority, it is omitted rather than hedged into the copy.

That is why the watermarking question does not worry us. A watermark, if one exists, indicates an AI tool was used somewhere in the process. It says nothing about whether the content is accurate, useful, or worth citing. Those are the things Google has said for three years that it actually measures.

What should your firm do next?

If your content program is producing posts that could run on any firm’s website in any state, watermarking is the least of your problems. That content was already invisible to AI search, and it was invisible before anyone started watermarking anything.

The fix is not more volume. It is content built around the questions your prospective clients are actually asking, answered directly, verified against primary sources, and specific enough that no national template could have produced it.

Too Darn Loud Marketing builds that kind of content program for law firms. If you want to learn more about what we do, contact us today to set up a call and discuss your specific marketing needs and goals.  

Frequently Asked Questions

Can Google tell if my law firm’s blog was written by AI?

For images, yes. Google can read SynthID and C2PA credentials and now surfaces that information to users in Search. For text, no reliable public detection exists as of yet. Third-party AI detectors produce frequent false positives and are not used by Google as a ranking input.

Does the watermark prove AI wrote my content?

No. A watermark indicates text was processed by a particular AI model. It cannot show how much a human edited afterward, and it does not survive substantial rewriting. Its absence proves nothing either, since it could come from an unwatermarked model or a human writer.

Will AI Overviews reduce traffic to my firm’s website?

Click-through rates on queries that trigger AI summaries have declined measurably, and Pew Research Center found users click a traditional result far less often when an AI summary appears. Citation still carries value though, and brands cited inside AI answers earn substantially more clicks per impression than uncited competitors on the same queries.

How long should a law firm blog post be?

Length is not the primary driver. Ahrefs analyzed pages cited in AI Overviews and found essentially no correlation between word count and citation. Structure and specificity matter far more than volume. The best rule of thumb is to use the word count needed to fully explain the subject matter. For most law firm blog posts, we have found that the ideal length is around 2000 to 2500 words.