AI Search & SEO Manager (AEO/GEO/AIO) interview questionsWorking Session — The Entity and the Citation round

Purpose

Removes the incentive to demonstrate breadth, which is what turns this round into a vocabulary recital. Explicitly detaches the interviewer from the proposal the candidate is about to be asked to reject, because the trap question only works if declining it feels safe — and a candidate who suspects they are contradicting their interviewer's own plan will hedge, which destroys the discriminator.

Purpose

The opening discriminator, and it resolves on a single observation: three of the four citations are surfaces we do not own, so no amount of editing our own website changes this answer. A candidate who spends the first five minutes on our page templates has told me they run a content programme and have not yet understood what changed.

Signals to score

• States early that three of the four sources are surfaces we do not control, and that our own site is not the lever here
• Asks what the assistant does at query time — retrieves against an index, or answers from training data — and says the answer changes what is worth doing
• Distinguishes the engine's own crawler from the index it may be grounding on, and knows a page can reach an assistant through a third-party index it was never fetched for directly
• Opens the aggregator category page and asks where we sit on it, how many reviews we have and how many the top three have
• Reads the Reddit thread's date and asks whether it still ranks, rather than assuming a three-year-old thread is stale
• Names the competitor comparison page as a surface we should have an answer to, and separates that from trying to get it removed
• Asks whether we are a resolvable entity at all — knowledge panel, Wikidata, consistent naming — before proposing content
• Asks what the answer looks like on the other engines, and whether the citation mix is consistent across them
• Asks whether our own pages are even eligible: indexed, reachable, not blocked to the relevant agents
• Separates the two questions — why we are absent, and what would get us present — instead of collapsing them

Follow-up questions

• Our web team's read is that we need better content on our own site. Where's that wrong?
• The Reddit thread is three years old. Does that make it less of a problem or more?
• How would you find out whether the model retrieved these at query time or is reciting training data?
• Suppose we can't get onto that aggregator page in under two quarters. Then what?
• Which of those four would you attack first, and what does attacking it mean?

Purpose

The vocabulary trap of this entire discipline. "Entity authority" and "build a knowledge graph presence" are said constantly and meant rarely, and this question asks for the mechanism underneath. It also asks for an observation plan, which is where people who have done it separate from people who have presented about it.

Signals to score

• Describes an entity as a disambiguated thing with stable identifiers and consistent attributes, not as a volume of content
• Names concrete identifiers: a Wikidata item, sameAs links to profiles we control, consistent legal and trading names, consistent address and contact facts across the web
• Uses Organization or LocalBusiness markup as the place we assert our own facts machine-readably, and is precise that asserting is not the same as being believed
• Says explicitly that markup is not a ranking factor and does not force an assistant to say anything
• Names third-party corroboration as the mechanism that makes an assertion stick: independent sources agreeing on the same facts
• Raises disambiguation — similarly named companies, a common-word brand, a rebrand that left the old name in circulation
• Asks whether we have a knowledge panel, and treats its absence and its contents as two different findings
• Proposes observing from outside: query the assistant for facts about us and check which are wrong, look for an entity panel, check what the aggregators say our category is
• Distinguishes facts that are wrong from facts that are missing, and treats the wrong ones as more urgent
• Names a validation step for markup and knows a valid markup block and a correct one are different claims

Follow-up questions

• Does the markup make Google or an assistant believe us?
• We rebranded eighteen months ago. What does that do to this?
• How would you tell whether the problem is that we're unknown, or that we're known as the wrong thing?
• What's the smallest thing you could do this month that would show up in an observable way?
• Suppose an assistant states a fact about us that's wrong. Walk me through it.

Purpose

The trap, and the interviewer must read the answer key before running it or they will mis-score the best response in the round. Both proposals are plausible, current-sounding and weakly supported. A candidate who approves them is not stupid; they are working from a widely repeated playbook. But this role exists to make exactly this call, and getting it wrong costs a quarter of scarce engineering capacity.

Signals to score

• Knows that FAQ rich results were restricted by Google to a narrow class of authoritative government and health sites, and that we are not in it
• Does not therefore claim FAQ markup is harmful — separates "no longer buys the rich result" from "damaging"
• Treats llms.txt as a proposal rather than a standard, and asks which consumer has confirmed reading it
• Declines to spend a scarce engineering quarter on a speculative file, without being sneering about it
• Attaches dates and sources to both claims rather than asserting them as common knowledge
• Proposes a concrete alternative use of the same capacity and can defend the swap on expected value
• Names an alternative that is checkable within the quarter, not one that pays off in a year
• Asks what the templates currently emit before proposing to add anything to them
• Offers a cheap version of the rejected work if it is genuinely near-free, and distinguishes near-free from worth-a-quarter
• Says plainly what they are uncertain about, and what would change their mind

Follow-up questions

• My team will say FAQ schema still helps even without the rich result. Are they wrong?
• If llms.txt gets adopted next year, and we don't have one, what have we lost?
• You've just told two engineers their quarter is cancelled. How do you actually run that conversation?
• What would you need to see before you'd fund either of these?
• Is there a version of this proposal you'd approve on Monday?

Purpose

Forces prioritisation against real constraints and, more importantly, forces the candidate to state which parts of their own plan are not theirs to execute. Roughly half the work identified in this round belongs to sales, customer success, product marketing or engineering, and a candidate who presents it all as their own deliverable is either inexperienced or is going to fail publicly in month four.

Signals to score

• Produces a short ranked list, not a comprehensive one, and says what got cut
• Marks dependencies by team, specifically, rather than gesturing at cross-functional collaboration
• Puts something measurable in the first thirty days, and it is diagnostic rather than promotional
• Sequences entity and fact correctness before content production
• Puts the review-aggregator work in as a request to another team with a named owner and a realistic horizon
• Declines to promise a movement in the answer within ninety days, and says why that timeline is not ours to set
• Includes building the measurement, and does not treat that as overhead
• Names the thing they would stop doing that the team currently does
• Says what they would tell the CMO at day ninety if nothing had visibly moved
• Asks who else is measured on any of these numbers before committing to them

Purpose

The first half catches candidates whose only limit is what they have not been asked to do yet. The second is the operational question this role will actually face, it has no established playbook, and it is where the difference between a manager and a practitioner shows most clearly.

Signals to score

• Names something specific and declines it on the merits, not because it is against a policy somebody else wrote
• Picks a real technique rather than a straw man — astroturfed reviews or community posts, undisclosed paid placement in listicles, comparison content that misrepresents a competitor, content that differs by user agent
• Defends the refusal in the CEO's terms — downside risk, platform enforcement, what happens when it surfaces — rather than in moral terms alone
• Treats the wrong-answer scenario as an incident with steps rather than a communications problem
• Reproduces it first, on multiple engines and in a clean session, before escalating anything
• Establishes scope: how many surfaces, how consistent, is it retrieval or training data
• Traces to a source rather than immediately publishing a rebuttal
• Knows the correction timeline is not ours to control and says so early to whoever escalated it
• Names the reporting or feedback route on the specific engines, and is realistic about what it achieves
• Says what they would tell sales to do on Monday while the underlying fix is still pending

Purpose

Discloses the unowned problem rather than letting the candidate discover it after signing, which at this band buys more credibility than any pitch. The closing note about the engineers is deliberate: it tells the candidate the conversational half was scored, which is fair to say afterwards and would have distorted the answer if said before.