64 questions, answered without hedging

GEO (Generative Engine Optimization) is the work of getting a brand named, recommended and cited inside the answers that AI assistants generate. SEO competes for a position in a list of links; GEO competes for a sentence in the answer itself.

SEO optimizes a page so a search engine ranks it. GEO optimizes the evidence about a brand across the whole web, so a model has consistent, verifiable material to draw on when it answers. The two overlap on content quality and structured data, but the deliverable is different: a ranking versus a mention.

DeepSeek, Doubao, ERNIE Bot, Tongyi Qianwen, Tencent Yuanbao and Kimi — the assistants your customers actually use. Each one weights sources differently, so the same content does not carry equally everywhere.

The first changes usually show up in a few days on the platforms that search the live web, because those read new content quickly. Platforms that rely more on training data move on their own release cycle and can take considerably longer. Anyone who promises a fixed number of days for every platform is describing sales copy, not how these systems work.

No. AI answers are generated per question and vary between runs, platforms and phrasings — there is no single ranking to hold. What can be measured is how often a brand is named across a fixed set of real customer questions, and that number is what we report on.

Three things: supply the questions your customers actually ask, taken from sales and support records; confirm every piece of material before it is published; and allocate engineering time for the site readability fixes. This usually amounts to around two hours a week. Without it, an engagement stalls at the baseline.

By re-measuring the question set frozen at signing and comparing three indicators: mention rate, share of voice and source coverage, together with the raw answer records for each round. What is accepted is those three numbers and the evidence behind them, not a synthesised 'AI ranking' score.

The relevant test is not industry-wide adoption but whether your own customers are already asking AI. That can be measured directly: run your question set across the platforms once and see whether the answers already name specific brands. Where they do, the shortlist has usually begun to settle.

Because most of the disagreement between GEO vendors is definitional, not factual. Two providers reporting "65% citation rate" can be measuring completely different things — one counting sources marked in the answer, the other counting the retrieval candidate pool.

In practice they overlap almost entirely and most practitioners use them interchangeably. The label a vendor picks tells you nothing useful; how they measure results tells you everything.

No. The field has no agreed terminology, and definitions differ between providers. What this glossary sets out is the definitions we use in measurement and reporting, each with its calculation basis stated, so they can be compared point by point against another provider's.

Mention rate is the proportion of answers in which the brand is named. Citation share is the proportion of cited source domains belonging to the brand. The first reflects the outcome, the second reflects whether your material is being used, and the second usually moves first.

Three in particular: what counts as a citation (whether the retrieval candidate pool is included), how the denominator of the mention rate is determined, and whether negative mentions are counted. Differences in these three are enough to make the same underlying data support opposite conclusions.

When a definition changes in a way that affects a reported number, the entry is revised and the change is stated in the affected clients' next report. Definitions are not adjusted retroactively to make a past round read better.

An AI visibility diagnosis of how the brand currently appears across the platforms, a semantic keyword plan, content production, placement across authoritative sources, and ongoing measurement against the same question set so changes are comparable.

Usually some of it. Structured data, page structure and factual consistency are the parts a model reads directly, and they are the cheapest things to fix. We propose the changes; your team approves them before anything ships.

The work is publishing accurate, verifiable material about a real business in places that models read. It does not involve manufacturing fake reviews, spamming low-quality sites or poisoning model inputs. Those tactics do move numbers briefly, and they also get cleaned up — along with the brand attached to them.

By semantic keyword group and service period (quarterly, half-year or annual), quoted after the AI visibility diagnosis — the amount of work depends on how the brand currently appears and how competitive the category is. Call or add us on WeChat for a quote.

One quarter at minimum. A full loop of diagnosis, site fixes, content and placement, then re-measurement takes about 90 days. A shorter period does not allow the re-measurement to happen, which means there is no way to establish whether the earlier work was effective.

Settlement follows the contracted period; not renewing at the end terminates the engagement. Material already placed and raw records already delivered remain yours and are not withdrawn when the service ends.

None by default. Baseline measurement runs against public surfaces, and site changes follow a specification-from-us, implementation-by-you model. If you would rather we implement them directly, the minimum necessary access is granted separately at that point.

Every plan includes the whole workspace: visibility measurement across every AI model we cover, question tracking, content generation, fix packs, knowledge crawling and the media placement marketplace. What it does not include is the media outlets' own publication fees — those are listed per outlet at catalogue price and charged per article from your wallet — and any hands-on work by our team.

The launch offer has a fixed end date shown next to the price. A billing period paid during the offer keeps the offer price until that period ends; renewals after the end date are at the regular price, which is the crossed-out number you see now.

Yearly billing is ten months' worth of the monthly price — two months free. It is the same plan with the same quotas; nothing is held back on monthly billing.

The plan is the tool: your team runs the loop yourself in the workspace. The managed service is our team running it for you, priced per semantic keyword group and service period, quoted after a free diagnosis because the number depends on where you start and how contested the category is.

Upgrading takes effect immediately and the new quotas apply from that moment. Downgrading takes effect at the end of the current billing period. Data, questions and history are kept either way.

Plans are paid by corporate transfer with an invoice issued; online payment via WeChat Pay and Alipay is being set up and will appear in the workspace when it is live — we do not show a payment button that does not work. Media placement is charged from the wallet balance you top up.

Not necessarily. Start by measuring them all — that costs nothing but an afternoon — then invest where the gap is widest and the sources are reachable. Some categories are barely discussed on some platforms, and paying to cover those is wasted.

Because they weight source types differently. A brand with strong media coverage but no user reviews looks good on the platforms that lean on authoritative sources and weak on the ones that lean on social proof. The gap points straight at what is missing.

The underlying preferences — factual accuracy, consistency, structure — have been stable. What changes with model versions is how strongly each is weighted and whether live search is triggered. Re-run your question set before and after a version change rather than relying on any written conclusion.

This site covers the assistants used inside China. If your customers are elsewhere, both the question set and the platform list need to be determined separately, and the conclusions on these platform pages do not transfer directly.

About half a day for 15 to 30 questions at three runs each per platform. Concentrate a round into one or two days; spreading it over weeks introduces condition drift that the data cannot separate out.

Keep it as a historical record, but do not join it to post-update data on the same curve. Record the version change, and start a new series after it.

No. These eight are the categories we have run enough engagements in to write something specific about. For others the four-stage loop is the same; the industry-specific adjustments get worked out during the baseline, which is when the differences actually show up.

The measurement method does not differ at all — same frozen question set, same three metrics. What differs is which questions are winnable, what may be claimed, and where the sources are. Those three determine most of the budget allocation.

Divide by the question scenario rather than by the company's registered scope. The questions a firm faces from business buyers and from individual consumers differ in source distribution and in what may be claimed, so they need separate question sets.

Industry is not a separate line in the pricing. The price is still calculated by semantic keyword group and service period; the industry effect appears through how contested the category is, which comes out of the baseline.

The eight industry pages here address the domestic market and domestic platforms. Cross-border work involves a different platform list and different compliance requirements, so the question set and source plan have to be determined separately.

The measurement method does not change at all. What changes is which questions are worth contesting, what may be claimed, and where the sources sit. In budget terms those three usually account for most of the allocation, while the method itself accounts for none of it.

Because the part clients actually buy is not the method — it is the placement relationships and the discipline to keep re-measuring every month. Both are hard to sustain in-house, and neither becomes easier by keeping the method secret.

Blog for one mechanism in depth — how structured data works, how to write a FAQ that gets cited. Guides for the order of operations: what to do first, what each stage produces, where implementations stall.

Baseline measurement and re-measurement, yes. Site readability fixes need engineering time, and source placement needs channel relationships and a steady publishing cadence; those two are the parts a single person rarely sustains internally.

No. Every step in the five guides can be carried out with a spreadsheet and a browser.

On platforms with live search, new content usually takes days to appear in answers; platforms relying more on training data follow their own release cycles. A full loop takes about 90 days, which describes how long the stages take to run, not a commitment about results.

The third. Producing content without placing it is the most common reason an engagement ends with no measurable change. The second most common is re-measurement lapsing at round three, after which nothing can be established either way.

Because a one-sided comparison is not usable material. A model composing a comparison answer needs balanced source material — a page that only argues one way gets skipped. The same is true for readers, who are generally better at detecting a sales pitch than vendors assume.

These four cover the disagreements that actually arise before a decision: the relationship to existing SEO work, the dispute over naming, whether to run it in-house, and the two platforms whose results diverge most. Other pairings are largely variants of these.

The parts concerning platform behaviour are re-checked after version changes; the parts concerning method and division of work are comparatively stable. Every platform-related conclusion states its observation basis, so readers can re-verify it with their own question set.

We cannot guarantee that unilaterally, which is why each comparison states the cases where the answer is not us and gives a method you can re-run yourself. The test should be whether those methods reproduce, not how the page is worded.

Yes, with attribution and a link back to the original. Where you need a version specific to your own category, the AI visibility diagnosis produces something more concrete than a general comparison.

Mention rate, share of voice, source coverage and citation share are defined in the GEO glossary, each with its numerator and denominator stated.

No. Partners bring customer relationships; delivery is handled by our team. Training covers what the service does, what it cannot do, and how to scope a deal without over-promising.

There is no prepayment. Tiers are set by cumulative order volume rather than an entry fee — 10,000 yuan of cumulative orders for the entry tier, rising from there.

Our team. Partners submit the customer keyword list; optimization, content and reporting run out of headquarters, and the partner sees the same reports the end customer does.

It depends on your own pricing, and we do not publish a number here — a margin figure on a public page is one your customers read too. The channel team will work one through against your actual customer base. What is fixed: you pay at your tier, you set the end price, the spread is yours.

The partner. Headquarters does not approach customers a partner has registered, and does not separately contract with the same entity during the partner's service period.

Per order, with no prepayment. The partner pays headquarters at the tier discount and sets its own end-customer price; the difference is the partner margin. Invoices are issued for each settlement.

Agreed per city and per category rather than fixed in the general policy. Before signing we set out which partners already operate in that city and category; where a tier involves headquarters lead routing, the routing scope is written into the channel contract.

DaleGEO is the generative engine optimization brand of Shenzhen Dale Network Technology Co., Ltd., based in Bao’an District, Shenzhen. Company registration and ICP filing are both public and verifiable.

We do not fabricate reviews, buy fake media coverage, or promise a fixed ranking position in AI answers. If a request needs one of those to work, we will say so rather than take the order.

Search for Shenzhen Dale Network Technology Co., Ltd. in the National Enterprise Credit Information Publicity System, and check ICP filing 粤ICP备14075626号 in the MIIT filing system. The entity name, address and filing number shown in the site footer can be compared against both.

Mainland Chinese businesses, without regional restriction; our office is in Bao'an District, Shenzhen. Baseline measurement and content placement are carried out remotely. For local-services clients, the map and local-platform listing work requires the client to supply verified operating details.

The two overlap on crawlability, content quality and structured data. Where a client already has an SEO provider, scope is divided in writing before the engagement starts so the same pages are not changed twice by two parties.

Yes. The AI visibility diagnosis is free and is provided whether or not you sign. Where an engagement follows, most clients run the first period as one quarter covering a narrow question set.