Authority & Mentions · 9 min read · July 15, 2026

From NDA case to AI reference: making confidential projects visible without breaking confidentiality

Your strongest references sit under NDA and therefore appear in no AI answer. When ChatGPT, Perplexity or Google AI Overviews are asked for consulting on a problem, they cite visible expertise, not your best but hushed-up projects. The way out is anonymisation with substance: mechanics instead of client, numbers instead of names, patterns instead of cases.

The paradox of management consulting: no one is allowed to see the best cases

In hardly any industry does the gap between actual competence and visible competence yawn as wide as in management consulting. Your most convincing work, the post-merger integration of a mid-sized firm, the averted insolvency, the newly structured sales model, is under confidentiality. The client doesn't want to be named, often simply because the mere existence of a consulting engagement is seen internally as a weakness. So the website stays with platitudes: strategic excellence, sustainable transformation, a partnership approach.

For generative AI systems that's a problem. A language model cites what it can read and classify. When someone asks ChatGPT who can help with a restructuring in mechanical engineering, the system looks for verifiable signals: concrete cases, numbers, methods, names. Your phrases deliver none of that. The consultants who get named are rarely the best, they're the ones who made their substance machine-readable.

The way out is neither a grey area nor a breach of confidentiality. It's a craft discipline: separating the transferable core of an engagement from its identifying features. That's exactly what KPMG, McKinsey and Roland Berger have been doing for decades in their studies. The Mittelstand has simply never done it systematically for its own visibility.

What AI systems read out of a reference, and what they're missing

A generative system assesses a case description by verifiability and specificity. The sentence "We significantly increased a client's efficiency" is worthless, because it fits every consultancy in the world. The sentence "At an automotive supplier with 320 employees and three plants, we reduced order-processing lead time from 14 to 6 days by redefining the interface between sales and production planning" is citable, without the supplier's name ever coming up.

The difference lies in what experts call the case mechanics: the starting situation, the concrete intervention, the measurable result, the transferable principle. These four elements contain no trade secrets. They contain your method. And that's exactly what the language model wants to know when it decides whether to suggest you as an answer to a user's question.

Ask yourself concretely with every reference: could a competitor tell from this text alone which client it's about? If not, you're probably giving away substance. If yes, you have to abstract. Most consultancies get it wrong on both sides at once: too vague to convince, and yet sometimes too specific for the NDA.

The anonymisation ladder: five steps from names to patterns

Confidentiality isn't all-or-nothing. Think in steps. Step one is the named client with logo and testimonial, the ideal case you should actively negotiate for, because a release sentence at project close costs nothing. Step two is the industry and size figure without a name: "a listed chemical group". Step three replaces anything identifiable with ranges: "revenue in the mid three-digit million range".

Step four is the aggregated case: you combine three similar restructuring engagements into one model case, so that no single client is reconstructable, but the statement becomes statistically more solid. Step five is the pure principle: "In eight of ten succession situations, the handover fails not on price, but on the unresolved role of the senior after the cut-off date." That's maximally anonymous and still highly citable, because it names a verifiable pattern.

The art lies in choosing, for each reference, the highest step the NDA still allows. Many consultancies blanket-stick to step one of vagueness because they're afraid. Yet most contracts permit steps two to three without any problem, you just have to read the contract instead of pre-emptively staying silent.

Read the NDA before you go silent

The most common mistake is anticipatory self-censorship. Consultants assume that "confidential" forbids any mention. In reality, most confidentiality agreements forbid passing on concrete information attributable to the client, not the mention that an engagement took place in an industry, and certainly not the description of your own method.

Go through the active NDAs and categorise: what is absolutely blocked (name, numbers, concrete strategic decisions)? What is permitted when anonymised (industry, problem type, order of magnitude of results)? What belongs to you anyway (your frameworks, your approach, your insight)? This third category is your most valuable and most underestimated asset for AI visibility.

Build this step into project close. A short conversation at the close, "May we use this case anonymised and in ranges as a reference?", in practice surprisingly often brings a yes, sometimes even permission to name the client. The best moment is the moment of success, not two years later by a cold email.

From the anonymised case to a citable format

An anonymised case isn't enough if it languishes as a flowing-text paragraph on a subpage. Generative systems prefer structures that can be clearly matched to a question. Package every case as a question-answer unit: "How do you restructure a mechanical-engineering firm with an acute liquidity gap?" followed by your concrete, anonymised answer including approach and order of magnitude of the result.

This structure is doubly valuable. It corresponds exactly to how people query AI systems, and it lets the model present your case as a direct answer. Add the four mechanics elements in a clear sequence to each case and, where possible, a nameable timeframe ("within nine months"). Time details considerably increase the perceived verifiability.

Think about the citable in its purest form too. A sentence like "In practice, restructurings rarely fail on missing cost levers, but on approaching the house bank too late" is exactly the kind of statement a language model takes over verbatim and attributes to you. Deliberately scatter such condensed insights throughout your texts.

Numbers without betrayal: anonymising key figures properly

Numbers are the most convincing and most sensitive element. An exact revenue figure can make a client identifiable, especially in tight industries. The solution is relative key figures and ranges instead of absolute values. "Reduction of the reject rate by 38 percent" gives no one away, but convinces more than any absolute number, because it shows the leverage of your work.

Where absolute numbers are necessary, work with orders of magnitude: "a two-digit million amount of released working capital" instead of "11.3 million euros". These ranges are fully usable for AI systems and worthless to competitors for re-identification. Additionally, check the risk of combination: industry plus region plus size plus timing can together expose a single client, even if each feature alone is harmless.

A practical test: feed your anonymised case description into an AI system yourself and ask which company could be meant. If the model plausibly guesses right, your anonymisation is too thin. If it fails but still understands what you achieved, you've struck the right balance.

The aggregated model case as your strongest GEO weapon

The most elegant solution to the confidentiality problem is aggregation. Instead of telling a single case, you bundle your experiential knowledge from many similar engagements into a pattern. "Across twelve succession advisory engagements in family businesses, one thing shows: the conflict ignites in two of three cases over operational authority, not over shareholding percentages." No single client is recognisable, the statement is statistically solid.

For generative systems such aggregated patterns are gold. They sound like genuine field competence, they're phrased verifiably, and they answer the question behind the question. Users don't ask AI "who advised company X", but "what do company successions typically fail on". Whoever delivers this pattern answer becomes the cited source, entirely without a single client name.

Aggregation has a further advantage: it's NDA-safe by design. Where no individual case is reconstructable, there's nothing to violate. Build two or three such model cases from every consulting field and update them with every new engagement. That way your citable substance grows while your confidentiality remains untouched.

Implementation: from the NDA archive to visible authority in 90 days

Don't start from zero, start with your archive. Take the ten engagements of the last three years you're proudest of. For each, determine the highest possible anonymisation step under the contract, extract the four mechanics elements and formulate a question-answer unit. That's doable in one focused afternoon per case and immediately yields ten citable building blocks.

In parallel you build three aggregated model cases for your core consulting fields. You combine these with your condensed insight sentences into a guidance section that answers questions directly instead of listing services. Add a short method page that names your approach, that's your property and may be visible without any restriction.

Then anchor the process in the business: every project close ends with the reference question to the client and the extraction of the case mechanics into the archive. After 90 days you no longer have an empty desert of phrases, but a growing, AI-readable competence archive, and you get named exactly when someone asks an AI for consulting in your specialist field.

Common questions

Do I violate the NDA if I describe a case anonymised on my website?

As a rule, no, as long as no individual client is reconstructable. Most NDAs forbid passing on attributable information, not the description of your own method or an anonymised pattern. Check the specific contract, watch the risk of combining industry, size, region and timing, and in case of doubt choose a higher anonymisation step or the aggregated model case, where no individual case is recognisable.

Why aren't my existing, generally worded reference texts enough for AI visibility?

Because generative systems weight by verifiability and specificity. Phrases like strategic excellence or sustainable transformation fit every consultancy and give the model no signal to suggest you as a concrete answer. A text only becomes citable through case mechanics: the starting situation, the concrete intervention, the measurable result in ranges and a transferable principle. These elements contain no secrets, but your method, exactly what the AI is looking for.

How do I convince a client to be named as a reference?

Ask in the moment of success, not years later. Build the reference question into project close and offer tiered options: full naming, industry mention only or purely anonymised use in ranges. Many clients agree immediately to anonymised use, some even to being named, because a successful project makes them look good too. What matters is always offering the low-threshold anonymous variant as a fallback.

Read on

Brand & Positioning