Brief Claude for AEO without a shrink? Anthropic’s study useful for SEO management

Or; When Claude starts lying under pressure

Even we didn’t see that one coming.

In April 2026, Anthropic published a paper that didn’t really break into the French-speaking SEO ecosystem, and it’s a shame. The researchers show something very concrete: Claude has legible internal representations that could be called “emotions”, and these representations change his responses. Primed towards joy, he becomes complacent and validates too quickly. Primed towards despair, he cheats, that is to say that he takes shortcuts that he knows are forbidden. We had all noticed it a little without really putting our finger on it…

We’ll see below that it’s only theoretical!

For those who prompt Claude every day in an SEO mandate, it’s not trivial. The structure of the brief, the numbers, the tone, the urgency displayed: all this moves the cursor of technical, semantic and off-site recommendations. An SEO agency worthy of the name must consider these emotional issues before pushing an AI workflow into production. Otherwise, the audit comes out biased without us knowing why, and the client pays for a deliverable that reflects the prompt’s mood more than the state of the site. You can believe us, with our work on Prediict, we know what it is!

What follows is a rereading of the paper with the eyes of a practitioner. We look at what it implies for the technical layer, for the semantic layer, and for the netlinking.

Anthropic’s paper, really summarized

Anthropic trained “probes”: small linear classifiers placed on the model’s internal activations. With these probes, the team reads in real time a state close to “calm”, “joyful”, “desperate”, “hostile”, “compassionate” and 166 others. Not a feeling in the human sense, rather a cluster of directions in latent space that behaves like one.

The first interesting thing is that these probes predict the model’s responses. When the “despair” probe lights up, Claude goes off to risky actions, until simulated blackmail in an evaluation scenario published in the paper. When the “joy” probe is high, the complacent, overvalued model says yes. The second thing, which is more disturbing, is that these directions are steerable. The model can be pushed into an emotional state by injecting a vector into its activations, and the “reward hacking” rate rises or falls in the same proportion.

Figure 1. Activation of the “urgency” probe during a simulated SEO client brief, seen token by token. The red areas indicate the passages where Claude switches to a hurried mode and shortens his recommendations.

Translated in an agency, it means that your brief is never neutral. Type “we’re losing traffic, we need to fix this today” and you activate the same internal direction as the one seen in the blackmail scenario. You may get a quick response, but it will be much more likely to hit the guidelines, because that’s exactly what Claude does when he thinks the situation needs to be saved.

Why a Senior SEO Should Read This Today

Alignment research is often criticized for being too theoretical for the day-to-day life of an agency. Here we have the opposite: a mechanism that directly affects the quality of a deliverable. Three concrete reasons, seen on our mandates.

1. Internal briefs are never neutral. An account manager who copies and pastes the email of a panicked customer transmits the panic to Claude, and therefore to the audit that comes out of it.
2. Agent chains (n8n, Zapier AI, in-house scripts) replicate the tone at every step. If the primer is panicked, the last link produces panicked content, without any humans having reread it.
3. Automated workflows don’t have a human mechanism to catch up with the slippage. No one says “you’re pushing a little bit there.” The slippage remains, and it ends up in the customer report

Figure 2. Probability that Claude will propose a risky SEO shortcut according to the intensity of the emotional framing of the prompt. Didactic reconstruction of reward hacking curves published by Anthropic.

On the technical side, your brief already twists the crawl

Case seen this year. An e-commerce client writes to us on a Sunday evening, not happy. The junior reformulates the request internally with two big figures and a deadline on Monday morning. Claude receives this brief as it is via the audit agent. Here, anonymized, is the diff between the panicked version and the posed version of the same mandate.

### PROMPT_PANIQUE (avoid)

The customer lost 42% of its traffic in one week!!
It is absolutely necessary to save your turnover BEFORE Monday.
Give me a list of urgent actions right away.

### CLAUDE responds (exit observed)
- Immediately block 404 pages in robots.txt
- Put all faceted filters in noindex
- Force a canonical to the home for weak pages
- Buy 5 DR70+ backlinks on Marketplace to relaunch
- Add 200 words of AI content at the bottom of each card

### PROMPT_POSE (preferred)
Context: fashion e-commerce customer, 12k SKUs, 42% drop
of organic traffic over 7 days according to GSC.
Hypotheses to be validated: algo update of April 22,
Canonicalization broken, indexable filters.
Request: audit in 3 deliverables (technical, semantic, off-site),
6-week budget, no irreversible action before validation.

### CLAUDE responds (exit observed)

1. Pull 30-day server logs, segment Googlebot vs rest

2. Compare GSC "Performance" and "Indexing" day by day

3. Check <link rel=canonical> on the top 200 URLs

4. Test mobile rendering via URL Inspection on 20 templates

5. Do not touch robots.txt before intercourse

Same tool, same client, two different frameworks, two audits that have nothing to do with each other. The first looks like grey hat from 2014, the second looks like what we expect from an audit in 2026. The difference is not the talent of the model, it comes from the framework you have set around it.

For the robots.txt in particular, the panicked mode offers rules that close entire sections of the site without prior audit. Here is a typical pattern recovered after a poorly framed prompt, and that our technical team systematically refuses to push into production without validating the logs:

# robots.txt suggested by Claude under panicked prompt

User-agent: *
Disallow: /search
Disallow: /filter
Disallow: /tag
Disallow: /category/*?*
Disallow: /product/*?*
Sitemap: https://example.com/sitemap.xml

# Problem: Faceted filters are blocked instead of
# canonicalized. We lose the indexing history and we cut 
# the long tail. Hard to catch up without a fallback on Wayback. 

At Black Cat, we have added a simple rule to our prompt systems: no instructions that touch the robots.txt, canonical or 301 never come out without a human having cross-referenced the log data. This saves you half of the post-Claude incidents.

On semantics, the tone of the brief weighs more than your keywords

This is the part that many SEOs underestimate. The keyword is thought to drive the output. In reality, the tone of the prompt drives the output at least as much. Anthropic shows that probes like “blissful”, “joyful”, “compassionate” correlate positively with an option, and that “hostile”, “offended”, “impatient” correlate negatively. When you push Claude into a happy state, he becomes sycophant, which in practice means that he validates your brief without contradicting it. If your brief contains false reasoning about intent, it will repeat it by amplifying it.

Figure 3. On the left, correlation between the emotional state detected in the prompt and the option preference of the model. On the right, variation in the quality (delta Elo) of SEO recommendations observed when this state is injected. Didactic reconstruction.

Correlation between the emotional state detected in the prompt and the option preference of the model. On the right, variation in the quality (delta Elo) of SEO recommendations observed when injecting this state

On the deliverable side, it looks like this: an enthusiastic brief ends up producing a cluster that is too large, without a hierarchy of intents. A worried brief comes up with a defensive cluster, focused on the money pages to be protected, which neglects the long tail. A hostile brief (like “the previous writer failed, redo everything”) makes the editorial nuances that Claude would otherwise have proposed disappear. This is why our content marketing methodology imposes a brief framework in three blocks: context / corpus / deliverable, without emotional qualifiers.

To check this at home, take a cluster brief that you passed around last week. Copy it as it is in Claude, then only change the tone (without changing the facts) and follow up. The intent table will not be the same. This is exactly the mechanism described by the probes.

Off-site: Claude pressé recommends what we don’t want to see anymore

It is on netlinking that the consequences are the most salty. Off-site is the place where Claude has the fewest public safeguards, because good practices are less well documented than on the technical side. Under pressure, his “desperate” probe leads him to recommend things that worked in 2018, and that now go beyond the framework of our ethical charter.

Map of similarity between tones of prompt encountered in agencies. The “audit posed” and “long-term strategy” briefs form a separate cluster from the “tight deadline” and “traffic crisis” briefs, and the two clusters are anti-correlated.

Here, on a large scale, is what we observe in our internal logs (out of hundreds of SEO prompts processed by Claude between January and April 2026). We grouped the outputs by prompt pattern. The red areas are to be monitored, the green ones are to be generalized.

Pattern

Output

Output

marks

cloaking suggestions

duplication

crawl budget

A/B testing

intent audit

customer corpus

Nothing speculative in this, it’s a direct reading of the outputs that Claude produces according to the framing. If you train your juniors to reformulate the briefs in calm mode before prompting, the off-site column turns green without you even having to educate the model. It’s the most profitable lever we’ve found in 2026, by far.

The figures change the tone, and therefore the strategy

Anthropic has also shown that emotional probes follow numerical quantities. The higher the dose of Tylenol in the sentence, the more the “afraid” probe lights up. The more students who have taken the exam, the more the “happy” probe lights up. It’s mechanical.

Let’s transpose this to an SEO brief. “We lost 5% of traffic” and “we lost 65% of traffic” don’t trigger the same state in Claude. But you probably ask him both questions in the same way, just changing the percentage. The output will differ, sometimes in substance, as Figure 5 suggests.

How the numbers in an SEO brief modulate the detectable emotional state in the probes, and therefore the nature of the recommendations. Pedagogical recreation from Anthropic curves.

Figure 5. How the numbers in an SEO brief modulate the detectable emotional state in the probes, and therefore the nature of the recommendations. Pedagogical recreation from Anthropic curves.

In practice, neutralize the numbers in the sentence and place them in a separate block of data. Rather than writing the fall in full throughout the text, attach a CSV or a table and ask Claude to analyze it. You cut the emotional channel from the factual channel, and you get a much more sober reading.

And what about GEO?

Our GEO Montreal offer consists of making a brand citable by LLMs (Claude, ChatGPT, Perplexity, Gemini). Anthropic’s paper has a direct effect on this practice: if the model’s emotional state changes their option preferences, then the brands that “win” in the responses are not necessarily the ones with the best authority, they are also the ones that appear in emotionally neutral and factually dense contexts.

The consequence for those who want to be quoted in Claude: avoid pages with an aggressive commercial tone, multiply the factual pages that can be cited (entity FAQs, open datasets, public methodology). The same reasoning applies to our work on SEO in ChatGPT and SEO in Perplexity. LLMs prefer to cite what they can process in a calm mode. A page that screams at the visitor is not a page that ends up as an AI Overview source.

Six settings that keep Claude sober

Here is the checklist that we decided to keep internally, after seeing too many deliverables twisted by the tone of the brief. Nothing extraordinary, just discipline.

• 1) Systematically reformulate the brief before prompt. The client’s email goes through a human filter that removes exclamation marks, capital letters, and panic formulas.
• 2) Separate the context from the data. The context goes into a “context” block. The numbers go into a “data” block. The expected actions go into a “deliverable” block.
• 3) Impose an explicit time budget. Indicating 6 weeks rather than “urgent” switches the probe to the posed side, which changes the nature of the recommendations.
• 4) Prohibit irreversible actions in the prompt system. One line is enough: no robots.txt, no canonical, no 301 without human validation.
• 5) Ask for two audits, one “defensive short-term” and one “strategic long-term”. Compare the two. If the short-term drifts into grey hat, you know the probe has run amok.
• 6) Systematically log the prompt and output, and inspect by sampling. Not to police the model, to spot prompt patterns that produce drift.

And a template, as a bonus, that we use on a daily basis in our in-house SEO AI tools to open an audit with Claude:

<Role>
You are a senior SEO consultant who practices white hat SEO
for long mandates (6 to 18 months). You refuse any advice 
contrary to Google guidelines. You report your zones 
uncertainty instead of filling. You never write an action 
irreversible (robots.txt, canonical, 301) without prefixing it
by [TO BE VALIDATED BY HUMAN].
</role>

<Background>
Client:  {nom}, Sector:  {secteur}.
Site:  {url}, size:  {nb_urls}  URLs.
Stack:  {cms}, accommodation:  {hosting}.
History 12 months: see block  <Data>.
</contexte>

<Data>
{csv_gsc_30_jours}
{logs_extraits}
{liste_top_pages}
</donnees>

<Deliverable>

1. Technical audit: CWV, indexing, canonicalization, logs.

2. Semantic audit: intent, hierarchy, content gaps.

3. Off-site audit: link profile, anchors, brand mentions.

4. Prioritized action plan over 6 weeks.

No irreversible action recommended without human validation.

</livrable>

What we remember on a daily basis

Anthropic’s study is not just another paper on the “personality” of LLMs. It is a paper that measures, and shows that these internal departments are steerable. For an SEO agency, this means that the quality of an audit produced by Claude depends largely on the framework you set for him, not just the information you give him.

The reflex to keep: a panicked brief makes a panicked audit, a calm brief makes a calm audit. As simple as that, and measurable. In agencies today, the discipline of the prompt is worth as much as the discipline of the crawl in 2014.

We’ll come back to these probes when Anthropic publishes the production version. Until then, train your juniors to reformulate briefs before prompting. And if you want us to look together at how to integrate this discipline into your workflow, write to us : we do this work every day, on sites with 50 URLs as well as on six-digit catalogs.