San Francisco · AI search
AI search agency in San Francisco
Your buyers were the first people on earth to start asking an assistant for a shortlist instead of searching. We find out whether you are on it, using a protocol you are welcome to pull apart.
Get a free AI visibility baselineWhat is generative engine optimization?
Generative engine optimization is the work of getting your company named and cited when someone asks an AI assistant for a recommendation. Because assistants synthesise an answer rather than return a list, the target is not a rank — it is inclusion in a set of three or four names, decided by what the wider web corroborates about you.
What actually decides whether you get named
Five factors, in roughly the order they matter. You will notice none of them is a keyword — that is the substantive difference between this and SEO.
- Whether the crawler can read you at all
- AI crawlers execute far less JavaScript than GooglebotA client-rendered site that Google eventually indexes may be effectively invisible here. This is the most common single cause we find in the Bay Area.
- Whether your page answers directly
- Citation rates rise 15-40% with direct answers and fact-dense tablesAssistants extract rather than read. Copy that builds toward its point is skipped in favour of copy that opens with it.
- Whether independent sources agree
- Corroboration across sites you do not controlOne site claiming something is marketing. Several unrelated sources agreeing is evidence, and these systems are built to weight it that way.
- Whether your description is consistent
- The same company, the same category, everywhereDifferent positioning on your site, your directory listings and your press coverage fragments the entity and suppresses you. Cheap to fix, rarely looked at.
- Whether anyone has written about you
- Third-party coverage in the sources engines retrieve fromThe slowest lever and the most durable. It is also why a smaller competitor sometimes gets named ahead of you despite ranking below you in Google.
We measure all five before recommending anything, because in about a third of cases the first row is the whole problem.
What the work actually consists of
Four workstreams. The first is diagnostic, the second is fast, and the last two are where durable position comes from.
Build a baseline you can audit
A fixed prompt set drawn from how your buyers genuinely phrase the question, run across five engines with repeated passes. You get the raw runs, not a score we invented — including every competitor named ahead of you and how often.
Reproducible by youMake the site machine-readable
Server-render what matters, expose content in HTML rather than behind scripts, get structured data agreeing with the visible page. For a lot of Bay Area companies this alone moves the numbers, because the content was always good and simply could not be read.
Frequently the whole fixRestructure for extraction
Direct answers in opening sentences, comparison tables, explicit figures with sources attached, definitions that stand alone when lifted out of context. Mostly applied to pages you already have rather than to new ones.
Editing, not writingBuild corroboration
Consistent entity data everywhere you appear, plus genuine third-party coverage and original research worth citing. Slow, unglamorous, and the reason positions built this way survive engine updates that wipe out cleverer tactics.
The part that lasts
How the engagement runs
Baseline, with the method disclosed
We tell you the prompts, the engines, the number of passes and the dates. If you want to run it yourself afterwards to check our numbers, that is a feature rather than a problem.
Fix retrieval, then extraction
Crawlability and rendering first, because everything downstream depends on it. Then content restructuring on the pages closest to a buying decision.
Re-run on a schedule
Same prompts, same engines, same passes. Movement reported against the original baseline, including the prompts where nothing changed — which happens, and you should hear about those too.
This field is about two years old, the engines change behaviour without notice, and anyone presenting it as a solved discipline is overselling. What we can defend is the measurement: a documented protocol, disclosed prompts, repeated runs, and honest reporting of the cases where our work made no difference.
Questions
- Is this measurable, or is it guesswork?
- It is measurable, provided you accept that the underlying systems are non-deterministic. That is why we run each prompt repeatedly across five engines rather than once — a single pass genuinely tells you nothing. You get frequency of appearance against a fixed prompt set, which moves in ways you can track and reproduce.
- Why would our site be invisible to AI crawlers but fine in Google?
- Because Googlebot renders JavaScript and most AI crawlers do considerably less of it. A client-rendered application that Google indexes on a second pass may present as an almost empty page to a retrieval crawler. It is the most common cause we find here, and server-rendering the commercially important pages usually resolves it.
- Is placement in an assistant something you can commit to?
- No. There is no paid placement, no submission process, and no agency can promise it honestly. What we commit to is measuring your current position properly, working the factors that demonstrably influence it, and showing you movement against a baseline you can verify independently.
- Does this replace our SEO programme?
- No, it sits alongside it and shares most of the technical foundation. The divergence is in content: SEO rewards depth and coverage, while this rewards clarity and extractability. A page can rank well and still never be cited, usually because it takes four paragraphs to say what it could have said in the first sentence.
- How much of this can we do ourselves?
- A fair amount, and we will tell you which parts. Entity consistency and answer-first rewriting are genuinely doable in-house if you have the time. What is harder to run yourself is disciplined measurement across five engines at sufficient repetition to be meaningful, which is where most internal attempts quietly stall.
Sources
- Oneskai GEO testing methodology — 100 fixed commercial prompts across 5 engines with 5 repeated runs (claim registry CLAIM-003).
- Generative Engine Optimization benchmark and RAG retrieval research — the 15-40% citation-rate finding (CLAIM-002).
- Google Search Central documentation on JavaScript rendering and crawlability.
- Oneskai AI Search Visibility Index — published engine-level visibility data.
Related
Find out if AI recommends you
A free baseline: how often you are named across five engines for the prompts your buyers actually use, which competitors appear instead, and the raw runs behind both numbers.
Method disclosed. Run it yourself afterwards if you want to check.