Claude's Watermarks Just Broke SEO — Transcript

Full transcript

1. 0:00Anthropic just started watermarking
2. 0:01every single word that Claude writes,
3. 0:03invisible, embedded in the text itself,
4. 0:05and it rides along when you paste it
5. 0:07onto your website. So, if you've been
6. 0:08publishing AI content, there's now a
7. 0:10permanent mark on it saying where it
8. 0:12came from. It's on everything, every
9. 0:15single model launched since August 2nd
10. 0:17across the API, Claude Code, Cowork, all
11. 0:19of it, worldwide. It survives copy and
12. 0:22paste. It survives editing. And
13. 0:24Anthropic says a detector is coming.
14. 0:26Plus, they're working on retrofitting
15. 0:28the older models. So, follow that chain
16. 0:30out. That mark is on your pages.
17. 0:33Anthropic is going to ship the detector,
18. 0:35and now Google knows exactly which pages
19. 0:38on the internet were written by AI down
20. 0:40to the exact paragraph. That's the fear,
21. 0:43not the watermark itself, but what a
22. 0:46search engine does once it can sort the
23. 0:48entire index into human and machine.
24. 0:52Now, if that flips into a demotion,
25. 0:54every AI published page you got turns
26. 0:56into a liability overnight. And there's
27. 0:58no real way to walk this back at scale.
28. 1:02That's the version currently circulating
29. 1:03right now, and every link in it sounds
30. 1:05reasonable. But none of this makes sense
31. 1:08until you understand what the watermark
32. 1:11actually is. Because almost everyone
33. 1:13reporting on it got this part wrong. The
34. 1:16watermark is not metadata. There's no
35. 1:18hidden file attached to your text.
36. 1:19There's no invisible characters buried
37. 1:22in it. The watermark is the word choices
38. 1:24themselves. When Claude writes, it's
39. 1:26choosing each word from a set of
40. 1:28options. The watermark nudges those
41. 1:30choices in a pattern. Every individual
42. 1:32word looks completely normal. But across
43. 1:35a few hundred words, the pattern of
44. 1:36choices adds up to a signature. This is
45. 1:39why it survives copy and paste. You're
46. 1:41still of course copying the words, and
47. 1:43the words are the watermark. It's also
48. 1:46why heavy rewriting kills it, but light
49. 1:48paraphrasing usually won't. You have to
50. 1:50change enough of the word choices so
51. 1:52that the pattern falls apart. Now, we
52. 1:54have to hold onto to because this
53. 1:56matters for something later. There's one
54. 1:58more thing that most people are getting
55. 2:00wrong here. A detected mark does not
56. 2:02prove that Claude wrote it, okay? This
57. 2:05is Anthropic's own language in their own
58. 2:08documents. If you wrote an entire
59. 2:10article yourself by hand, you paste it
60. 2:12in, and you ask Claude to clean up the
61. 2:14grammar, the output will still likely
62. 2:16carry that mark. It's the same if you
63. 2:19ask it to translate or summarize. All it
64. 2:22tells you then is that the text passed
65. 2:23through Claude at some point, which is a
66. 2:26much narrower claim than a lot of people
67. 2:28are making. So, that's what the mark is.
68. 2:31It's real, it's durable,
69. 2:34and it's a lot dumber than most of the
70. 2:35coverage suggests. But, the version
71. 2:37circulating right now breaks in three
72. 2:39separate places. I'm going to show you.
73. 2:41First,
74. 2:42Anthropic ships a detector, and it gets
75. 2:45into Google's hands. So, okay, this one
76. 2:48clearly plausible. Anthropic has said it
77. 2:50intends to publish the details and allow
78. 2:52third parties to check it. So, I'm
79. 2:55obviously not going to stand here and
80. 2:56tell you that detection is impossible.
81. 2:58It is likely coming. This makes the
82. 3:00second link the important one because
83. 3:02the whole chain assumes that once Google
84. 3:05can detect AI content, Google will do
85. 3:08something about it. And honestly, that
86. 3:10assumption has never been tested with
87. 3:13text. Text watermarking is like 10 days
88. 3:17old. Nobody has data on it. Nobody can
89. 3:20have data on it. But, it has been tested
90. 3:22somewhere else for 3 years in public,
91. 3:25and almost one in this industry has
92. 3:28really bothered to look. So, before I
93. 3:30show that to you, I want you to look at
94. 3:32what Google has actually said about
95. 3:34AI-generated content because their
96. 3:36documented position is not silence. They
97. 3:39wrote it down. Their guidance says
98. 3:41appropriate use of AI or automation is
99. 3:44not against their guidelines, and that
100. 3:47however content is produced, you should
101. 3:49be aiming for original, high-quality,
102. 3:52people-first work. This is Google's own
103. 3:55blog post from 2023 and it's still up.
104. 3:57Now, I'm not going to sit here and tell
105. 4:00you that you should trust Google's
106. 4:01public statements. Google has spent
107. 4:03years telling this industry that clicks
108. 4:06were too noisy to use in ranking. Then,
109. 4:09their VP of Search testified under oath
110. 4:11that Navboost is one of their important
111. 4:13signals and later a former engineer
112. 4:16testified that Google was told
113. 4:18internally not to admit this publicly
114. 4:20because SEOs would exploit it. This is
115. 4:23the same story with the helpful content
116. 4:25update. Google said it would reward
117. 4:27people-first content. Then, Danny
118. 4:29Sullivan sat in a room with publishers
119. 4:32it had crushed and told them there was
120. 4:34nothing wrong with their content.
121. 4:35Google's systems just weren't surfacing
122. 4:37it. Those people paid their own way to
123. 4:40that meeting. Okay, so a Google blog
124. 4:42post proves basically nothing on its
125. 4:44own. What I want is behavior, years of
126. 4:47it, with the detection already working.
127. 4:49This brings me back to what's been
128. 4:51running in public since 2023, images.
129. 4:54Google has a watermarking system called
130. 4:57SynthID. It's how Google marks the
131. 4:59images, the video, and audio its models
132. 5:02produce. And it's been live since 2023.
133. 5:05It shows up inside Google Search, in
134. 5:07Lens, in Circle to Search, where it'll
135. 5:10tell you whether a picture was
136. 5:12AI-generated. Now, every major lab does
137. 5:15some version of this. It's been the
138. 5:17industry standard for years. So, if
139. 5:20Google were ever going to rank against
140. 5:22AI content, images are where this would
141. 5:25have happened first. The mark exists,
142. 5:28the detection works, it's been sitting
143. 5:30there for over 3 years. Now, let me show
144. 5:33this to you. This is a page that my Core
145. 5:3530 agent designed for local SEO built.
146. 5:38And here's an image that was
147. 5:40AI-generated, start to finish. It's a
148. 5:42good image. Google Gemini generated it.
149. 5:45You'd scroll right past this if you
150. 5:46visited this website and never really
151. 5:48think twice. I mean, maybe it looks a
152. 5:50little weird that he's taking a wrench
153. 5:51to this part of the water heater, but
154. 5:54it's not a super weird image. Now, let
155. 5:56me show you what this looks like on a
156. 5:58free detector. This is just one random
157. 6:00detector I found and it says 99% AI
158. 6:04generated. This is off the pixels alone
159. 6:06in like 2 seconds. There's no login,
160. 6:08there's no key, there's no special
161. 6:10access. Anybody can run this detection
162. 6:12including Google and honestly Google can
163. 6:15do it better because Google's own model
164. 6:17made the image and Google's own
165. 6:19watermarks are in its output. So, this
166. 6:22is not a close call. Every layer of
167. 6:25detection that exists catches it. Now,
168. 6:27here's what happened to these pages. Let
169. 6:29me show you some rank maps. This is from
170. 6:31people running that very same Core 30
171. 6:33agent before and after. AI written
172. 6:35content, AI generated images published
173. 6:38to their sites. Here's Here's this one,
174. 6:40May 27th to June 16th. So, this is
175. 6:43roughly 3 weeks of a before and after
176. 6:46all based on content written by the Core
177. 6:4830 agent. Here's another one. Uh this
178. 6:51May 2nd to June 25th. So, this is almost
179. 6:532 months. Uh I can show another page
180. 6:56here.
181. 6:57Uh we have This is July 16th to July
182. 7:0025th. So, this is 9 days, right? Insane
183. 7:03progress for 9 days. We can show another
184. 7:05one here. This is the before, June 29th.
185. 7:08Here's the after, July 24th. A little
186. 7:11bit less than a month. Absolutely crazy
187. 7:14results that people are getting with the
188. 7:15Core 30 agent. And remember, this is
189. 7:17with AI generated images, AI generated
190. 7:19text, 3 years of working free public
191. 7:22detection and it has never once been
192. 7:24what decides whether a page ranks.
193. 7:27That's the second link gone. Google
194. 7:29having the capability and Google using
195. 7:31it against you have never been the same
196. 7:33thing and these are the receipts to
197. 7:36prove it. Which leaves the third link.
198. 7:38And this is the one that's going to cost
199. 7:40you money if you get it wrong. Because
200. 7:42somebody is going to tell you that you
201. 7:45need to strip the watermark out anyway
202. 7:47just to be safe. Now,
203. 7:50these methods exist. There's published
204. 7:52research on this. Documented academic
205. 7:54papers. You can run text through a
206. 7:56paraphraser. You can translate it into
207. 7:57another language and back. You can do
208. 7:59character level attacks that swap
209. 8:01letters for look-alikes from other
210. 8:03alphabets. And there's newer work on
211. 8:05more targeted removal. Some of it
212. 8:07actually needs access to the model's
213. 8:09internals, which obviously you're never
214. 8:11going to get through a hosted API, but
215. 8:13most of it doesn't. Most of it you can
216. 8:15just run this afternoon on content if
217. 8:17you wanted to. But they all cost
218. 8:19something. Remember how the watermark
219. 8:22works, that the signal lives in the word
220. 8:24choices themselves. That's why
221. 8:26paraphrases and round trip translation
222. 8:28defeat it. But keep in mind, you can't
223. 8:30use Claude or OpenAI or Gemini to do the
224. 8:33paraphrasing or to do the round trip
225. 8:35translation. This has to be done by
226. 8:37hand. OpenAI, Gemini are signing on to
227. 8:41their own watermarks. This is part of a
228. 8:43European Union law. So, we're not
229. 8:46deleting a tag. We're not turning every
230. 8:48word on the page. Doing this, especially
231. 8:51by hand, the writing will flatten,
232. 8:53errors will creep in that no one
233. 8:54proofreads for, and you've ended up
234. 8:56rewriting a page that was working. And
235. 8:59the character swaps are the easiest but
236. 9:01the worst of the bunch because of how
237. 9:04this works under the hood. Remember that
238. 9:06Google doesn't match keywords, it
239. 9:07resolves entities. It reads the city
240. 9:10Molden and connects it to everything we
241. 9:12know about that city. It reads your
242. 9:14business name and connects it to your
243. 9:15Google Business Profile. But none of it
244. 9:17happens visually. Of course, Google's
245. 9:20algorithm doesn't have eyes. Text gets
246. 9:22broken into tokens before anything looks
247. 9:24at it. And if you've replaced the Latin
248. 9:27symbol A with a Cyrillic character that
249. 9:29looks identical, it's a still a
250. 9:31completely different character
251. 9:33underneath. So, a swapped word, it's not
252. 9:35a misspelling that something can work
253. 9:37out from context. It doesn't resolve to
254. 9:40anything at all. Do that to a service
255. 9:42name, a category, a city, the business
256. 9:44name, and the page stops making the
257. 9:46connection it was built to make. On top
258. 9:48of that, mixed script text is trivially
259. 9:51detectable, and Google has run this as a
260. 9:53spam signal for years. You'd be breaking
261. 9:56the entities on your page. You'd be
262. 9:58handing Google something it flags on
263. 10:00site. You'd be degrading real pages to
264. 10:02hide from a consequence that doesn't
265. 10:05exist. Okay, so look, things change. If
266. 10:07there's ever evidence that Providence
267. 10:08affects ranking, I'll talk about it, and
268. 10:11I'll say so plainly. I have enough
269. 10:12clients under management that are
270. 10:14ranking with AI-written content and
271. 10:16AI-generated images that we would see it
272. 10:19immediately at my agency. But right now,
273. 10:21the evidence runs the other way, and
274. 10:23it's been running that way for 3 years.
275. 10:26If you want to see the system that
276. 10:27produced these rank maps, I put together
277. 10:29a full walk-through of how the Court 38
278. 10:31Agent works, what it writes, and what
279. 10:33the pages look like. You can check out
280. 10:35that video here.