A developer asks which practices have delivered occupied-building workplace retrofits in a given city. A procurement lead asks which architects have experience with accessible cultural buildings under public framework conditions.
A hospitality group asks who combines heritage conservation with hotel interiors. A private client asks who designs low-energy houses in a cold climate. A journalist asks who is doing interesting work in industrial adaptive reuse.
Five questions about the same discipline, often about the same practice. The answers rarely name the same firms, because each carries a different building type, constraint and notion of proof.
Architecture has adopted AI faster inside the studio than around it. RIBA’s 2026 AI Report found that nearly three-quarters of practices, 74 percent, now use AI on at least some of their projects, with 75 percent reporting an improvement in productivity and 57 percent a positive return on investment.
Far fewer have checked the other direction: what AI systems say about the practice when a prospective client asks who belongs on a shortlist.
That is the question generative engine optimization deals with. For a practice, the unit worth measuring turns out to be the client role, not the firm as a whole.
GEO is a research field, not a marketing coinage
The term comes from a paper by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, published at ACM SIGKDD 2024.
The authors built GEO-bench, a benchmark of user queries with the web sources needed to answer them, and tested nine ways of rewriting a source to see which changed how generative engines cited it.
Three findings matter for a practice.
- Evidence beat wording. Citing sources, adding relevant quotations and adding statistics produced the largest gains, roughly 30 to 40 percent relative improvement on the paper’s position-adjusted word count metric, with visibility gains of up to 40 percent overall.
- Keyword techniques failed. Stuffing terms into a page did not persuade a generative engine to cite it.
- Effectiveness varied by domain, which led the authors to argue for domain-specific optimisation. A playbook written for SaaS does not transfer, because the questions, constraints and authoritative sources differ in the built environment.
How an answer about your practice gets built
Google names two mechanisms in its Search Central guidance, and both explain why practices with strong work go missing.
The first is grounding: the core Search ranking systems retrieve current pages from the index, and the model writes its answer from that material rather than from memory. The second is query fan-out, concurrent related queries the system generates to gather more than the literal question would return.
Applied to a client question, one prompt runs roughly like this:
- The client asks a full sentence with constraints in it: typology, location, budget band, programme risk.
- The system generates sub-questions it was never given, one per constraint.
- Those run against the index, and only pages that are indexed and eligible to appear with a snippet can be retrieved.
- The model writes one answer from what came back, naming a few practices and attaching links.
- A follow-up reuses the conversation, so the second answer can rest on a partly different retrieval.
A practice can drop out at any step: at three because the relevant text lives inside a drawing, at four because no retrieved page states the constraint the client asked about, at five because the follow-up moved to procurement and nothing on the site covers it.
Why a single visibility score misleads a practice
Most AI visibility reporting produces one number: how often a brand appears across a set of monitored prompts. For a firm with one product and one buyer that is a reasonable summary. For an architecture practice it hides the only thing worth knowing.
Consider a mid-sized studio with a healthcare portfolio, a growing education arm and one well-published cultural building. A blended score might read as respectable.
Underneath it, the practice may be the first name AI gives for accessible cultural projects, invisible for healthcare planning where a competitor has published post-occupancy data, and marginal for education because its school projects sit in the archive as image galleries with no text about the brief.
Averaged together, those three positions produce a figure that moves up and down without telling anyone what to do about it. Kept apart, they produce a decision: defend the sector you own, push on the one you nearly own, and make a deliberate call on the one you are losing.
Writing prompts the way clients actually ask
Persona prompts are not keywords with question marks attached. A client’s question carries constraints, and those constraints are what the retrieval step keys on.
A usable prompt tends to contain four elements: the role asking, the building type, the place, and the condition that makes the project hard.
Compare a keyword-style prompt, “best architects adaptive reuse”, with a client-style one: which practices have converted listed industrial buildings into mixed-use schemes in the north of England while the site remained partly operational. The second returns a different set of firms, because it asks for evidence rather than reputation.
A practical persona set usually looks like this:
- The commercial developer, asking about programme risk, planning history and delivery at scale.
- The public-sector client, asking about framework experience, accessibility and procurement compliance.
- The specialist operator, a hospital trust, school group or hotel brand, asking about typology-specific performance.
- The private client, asking about process, budget bands and how the practice works.
- The journalist or awards juror, whose coverage the other four personas later rely on.
Each persona also has a language and a geography. The same question asked in German about a Berlin project set retrieves a different source pool and returns a different shortlist.
What the answer contains besides your name
Presence is the crudest measure. Four other readings carry more information for a practice:
- Position. First name in the shortlist, or the last one in a closing sentence?
- Framing. Practices are often described accurately but narrowly, as the housing office or the conservation office, when the strategy is to grow another sector.
- Competitor set. Which firms occupy the same answer, which is frequently not the peer group the studio believes it competes with.
- Citations. Where the description came from: an awards database, an architecture publication, a planning portal, a consultant’s project list, a Wikipedia entry.
The last one is the actionable reading. If the model’s account of your healthcare experience comes from a five-year-old competition write-up rather than your own project pages, the correction is partly editorial work on the site and partly keeping the cited sources accurate.
Turning spot checks into measurement
Manual testing is worth doing once and stops being informative quickly. Answers vary by wording, model, language, geography and date, so a screenshot cannot separate a real change from ordinary variance. A signal needs a stable prompt set, run repeatedly, with the full answers kept.
For a practice, the deciding capability is whether that measurement is organised by client role rather than as one blended average. A GEO Plattform which offers prompts based on buyer persona gives each persona its own prompt set, its own competitor list and its own score, so the developer question and the public-sector question never get averaged into one figure.
Truffle drafts those prompts from the practice’s own domain and personas, then runs them daily across six surfaces, ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews and Google AI Mode, recording mention rate, average position and link inclusion separately for each role.
The output is a status per persona instead of a trend line: the roles you own, the ones that are contested, and the ones a competitor holds. A persona and competitor grid shows which rival occupies which role, which is usually where a marketing conversation becomes a business-development conversation.
Citation tracking lists the domains and pages the answers lean on, including frequently cited sources where the practice does not appear at all, which doubles as an outreach list. Each finding comes with a prioritised recommendation across content, technical work and outreach.
The original answers stay stored with their model, date and wording, so a change can be traced instead of argued about. Personas are capped on entry plans and unlimited on the agency tier, with a seven-day trial that needs no credit card and a free single audit without an account.
One limit belongs in plain sight. No external tool has access to the internal ranking or retrieval systems of Google or any AI provider, a point Google itself makes in its guidance.
What a tool can do is ask the question, record the answer and turn observations into a trend. For a discipline this new, that is enough to work with.
What actually changes the answer
Measurement earns its cost only if it leads to work, and for a practice the work is mostly making existing knowledge legible.
Project pages carry the weight. A page that names the building and shows twelve photographs gives a retrieval system almost nothing.
A page that states the brief, the typology, the location and the conditions that shaped the design, the practice’s actual role, the constraints, the significant materials, measured outcomes where they exist, the certification with its date, and the collaborators gives it the evidence the GEO paper found to matter.
None of that requires abandoning the visual argument. It requires a short factual block and captions that explain what a drawing demonstrates instead of repeating the project title.
Sector and service pages turn single projects into demonstrated expertise, which is what a persona question asks for. One school is a project. Six schools plus a page on how the practice handles daylight, acoustics and phased construction on occupied sites is a claim a model can attribute.
The technical groundwork is dull and well documented. Google states that established SEO fundamentals still apply to its AI features, that a page must be indexed and eligible to appear with a snippet, and that no special AI file, markup or schema is required.
Its July 2026 guidance says llms.txt files, content chunking and rewriting text for AI systems do nothing for visibility in Google Search. What helps is text that exists as text rather than inside an image, crawlable pages, internal links, and structured data matching what a visitor sees.
Two cautions. Google warns that generating pages for every query variation, mainly to influence rankings or AI answers, falls under its scaled content abuse policy. The answer to a persona gap is one good page, not forty thin ones.
And the incentive to make claims explicit is also an incentive to overstate. Project status, the practice’s role, dates, collaborators and certifications have to be exact. A model that finds your page contradicting a client’s press release has no reason to prefer your version.
Two technical traps specific to architecture websites
Practice websites are built to look good, which puts them at the wrong end of two documented problems.
Most AI crawlers do not run JavaScript. The reference dataset is the Vercel and MERJ server-log analysis from December 2024, covering hundreds of millions of real crawler fetches.
It found no evidence of JavaScript execution by the major AI crawlers. GPTBot downloaded JavaScript files in roughly 11.5 percent of requests and ClaudeBot in roughly 23.8 percent, without running them.
Googlebot renders. So a portfolio can perform respectably in Google Search and still be an empty frame to ChatGPT, Claude or Perplexity.
The usual suspects on a practice site: project facts held in a slider or lightbox, text that loads only when a section scrolls into view, credits and data sheets inside PDFs, and a cookie banner standing between the crawler and the content.
The test takes a minute: fetch one project page without a browser and search the raw response for a sentence you can see on screen. If it is missing, the retrieval layer does not have it.
Access may already be switched off. Many sites blocked AI crawlers in 2024 and never revisited it. Two details are worth checking.
Google-Extended is a robots.txt token governing training and grounding in Gemini Apps. Google states it does not affect inclusion in Google Search, so disallowing it does nothing to AI Overviews, while it can affect how the Gemini app describes your practice.
On 1 July 2026 Cloudflare introduced categories for AI bots. From 15 September 2026, because the most restrictive applicable setting wins, it states that Googlebot, Applebot and BingBot will be blocked for customers who selected to block Training, including anyone who used its legacy one-click “Block AI bots” service.
That enforcement happens at the edge, so nothing in robots.txt will show it.
Getting the third-party record right
Independent studies through 2026 keep finding that a large share of AI citations goes to third-party platforms rather than the brand’s own site. The figures differ wildly between vendors and move week to week, so treat any single percentage with suspicion. The pattern holds regardless.
For a practice, the record outside your domain does a lot of the describing:
- Awards and competition databases. Often the most-cited evidence of a typology credential, and often the most out of date.
- Architecture publications. A project write-up carries the brief, the constraint and the outcome in text, which is exactly the shape retrieval prefers.
- The encyclopaedic layer. An outdated Wikipedia entry, practice directory listing or professional body profile keeps resurfacing in answers long after the website is updated.
- Consultant and contractor pages. Engineers and contractors list the same project from their side. Where those descriptions contradict yours, the model has no reason to prefer your version.
- Named specialists. Team pages and professional profiles connect a person to a typology, which is how a model justifies naming a firm for a specialist question.
One caution. Google’s guidance explicitly lists chasing inauthentic mentions among the things that do not work, so the aim is an accurate record in places that already cover the field, not manufactured coverage.
A workflow that fits a practice
- Pick one sector or service to grow, instead of auditing everything.
- Write down ten to fifteen questions a client in that sector asks before assembling a shortlist, in their words, constraints included.
- Split them by role, market and language, and record a baseline answer for each with its citations and competitors.
- Rewrite the project pages that are the strongest evidence for each gap: brief, context, role, constraints, outcomes.
- Connect them to the matching sector, service and team pages, and correct the third-party profiles that get cited.
- Re-run the same prompts over the following weeks and compare.
Frequently asked
Is this different from SEO for architects? The technical groundwork overlaps. The measurement does not. Rankings measure a page against a query. Persona-level GEO measures whether the practice is named in a synthesised answer, in what position, against which competitors, and on the strength of which sources.
How many personas does a practice need? Three to five to start. Developer, public-sector client, specialist operator, private client and the press or awards audience covers most practices, each with its own prompt set.
Do we need to publish performance data we do not have? No, and inventing it is worse than omitting it. Where measured outcomes exist, state them with method and date. Where they do not, describe the strategy and the constraint honestly. The GEO research rewards verifiable evidence, not confident adjectives.
The portfolio as a knowledge system
A portfolio will always be a visual argument, and no optimisation framework should flatten spatial experience into generic prose.
But a portfolio now has a second job: to explain the conditions behind the image, the decisions behind the form and the results behind the claim, in text that can be retrieved and attributed.
Practices that do this get something more useful than a score. They find out which of their five clients is being sent elsewhere, and which project in their own archive was the answer.
