Introduction
Search engines have been studied
from the point of view of the currency of the information in their database
indexes (Brewington & Cybenko, 2000), instabilities in the results
(Bar-Ilan, 1999, 2001; Bar-Ilan & Peritz, 2002), economical and language-based
inequalities in the search engine results (Introna & Nissenbaum, 2000;
Vaughan & Thelwall, 2004; Van Couvering, 2004), and the lack of
interactivity on the Web (Wouters & Gerbec, 2003). Most studies focus on
the performance of various search engines from the point of view of a general
user (Risvik & Michelsen, 2002; Lewandowski, 2004). Our focus is not on
general users nor on search engine performance but on the theoretical and
practical implications of search engine use for scholarly research. The way
search engines re-write the past by updating their indexes in the present has
hitherto received little attention (Wouters et al., 2004). In this
paper, we address a set of questions relating to how search engines can be
considered as ‘clocks’ of the internet that tick with different frequencies.
More specifically, we are interested in the way the updating affects the
present that is produced by search engines and in which they evolve.
The question of how temporal
representations change over time is an urgent one. In every social reality,
temporality is central to the network of relationships. Societies reconstruct
themselves by reconstructing also their histories. This can be considered as a
constant process of mutual adaptation between historical traditions and
institutions, and between emerging expectations about the future and
appreciations of the past (Schütz, 1932). The duration of activities and
processes, and the ways in which they are synchronized and updated, affect the
positions of agents in the network. The network development itself can be
considered as an interplay and interaction effect among the various
temporalities involved (Innis, 1952; Nowotny, 1994).
In terms of systems theory, this
can be understood as an interference among the updating frequencies of the
subsystems in society. The subsystem of science, for example, publishes
scientific results with a frequency very different from that of newspapers.
Similarly, some Web pages are updated with a frequency higher than others, and
different search engines update their indexes with structurally different
frequencies (Thelwall, 2001). Furthermore, new pages are continuously added to
the Web and old ones are removed from the Web. A focus on the different
updating frequencies and their temporality enables the analysis of
socio-technical systems in which technical constructs are functioning both as
nodes and as media facilitating relationships between the nodes of the network
(Latour, 1988; Leydesdorff, 1994, 2001).
The study of updating cycles has an
especially salient relevance to search engines. Some search engines (for
example, AltaVista and Google) can be used to search for
information in the internet for specific periods of time.
However, these ‘date stamps’ are not determined by the first occurrence of the
pages in the Web, but by the last date at which a page was updated or a new
page was added and the search engine’s crawler updated this change in the
database. For the update in the search engine database, any alteration of the
Web page may count as a change, no matter how minor it was. The ‘same’ Web page
may therefore belong to the year 1995 in a data set collected in 2003, while in
a data set collected in 2004 it belongs to the year 2003—or it may have been
‘forgotten’ by the search engine altogether (Bar-Ilan, 1999). Hence, when they
are used to search for historical dates, search engines represent the results
of the interacting frequencies of (a) the creation and updating of Web pages
and (b) the retrieval and updating at the level of search engine indices. The
results are not likely to reflect the dates of publication of the documents
under study. This has implications for the use of search engines in scholarly
research.
While the development
of the engines remains historical, their dynamics evolve in the present and
reflexively to the system to which they belong (that is, the internet). Thus,
these engines reconstruct their histories by looking backwards. In other words,
search engines provide the past with a ‘meaning’ and can thus be considered as
anticipatory systems (Rosen, 1985; Dubois, 1998; Leydesdorff, 2005). Because of
the updating effects, such reconstructions will tend to draw Web sites into the
most recent past, thereby possibly erasing the older representations of the
same Web pages. Search engines catalogue the Web, and these catalogues are
continuously updated in order to keep them current.
Research
Questions
In this study we
attempt to test how the three updating frequencies (updating the Web pages,
updating the search engine database, and the growth of the Web) resonate at the
internet. Search engine results allow us to study empirically the constant
change in the multiple presents. We compared two search engines by performing
searches with exactly the same search string at different moments of time. The
focus was on the two major search engines that provide the option to limit
searches to specific dates. AltaVista’s Advanced Search Engine (www.AltaVista.com/web/adv) allows
searches from the year 1980 to the present, limited to specific dates, months,
or years. Google is currently the most frequently used and largest
search engine (www.searchengineshowdown.com).
It provides the option for similar date-specific searches via Google’s
APIs or Faganfinder (www.faganfinder.com/Google.html).
The latter engine exploits the database of Google.
Originally, we planned
to provide search results with a one-year time interval (January 2003 versus
January 2004) and a one-month time-interval (January 2004 and February 2004).
During our study, however, AltaVista changed its search engine to the
one of Yahoo! (April 2004). The number of hits thereafter declined
considerably, and therefore we decided to conduct an additional search at the
end of April 2004. In general, search engines function very differently. The
exact algorithms used by the various engines are commercial secrets, but it is
known that while Google uses link-based crawling for updating its
database, Altavista relies on a keyword-based crawling (www.searchenginewatch.com).
We are interested in
two related questions. One is the question of the extent to which the same
results can be reproduced using search engines for searches at different
moments of time, i.e. at different ‘presents’. Because of the updating
mechanisms, one can no longer assume that time-series data reflect historical
developments of the systems under study. This raises the question whether one
can construct time series data by periodically searching the Web for specific
retrieval terms. To which extent can these results be reproduced? What does the
level of reproducibility reveal about the resonance between the various
updating frequencies?
The second question is
related: How can the changes in the results be interpreted? It seems too easy
to conclude that this type of data is worthless, since the ‘errors’ are
generated systematically. The updating mechanism represents a significant
socio-technical activity on the Web. At the same time, the updating of the Web
pages provides us with an empirical domain to study this mechanism of change.
What kind of windows on the reality of the Web do the search engines provide?
Before addressing the
technical details of the experiment, let us first specify our theoretical
expectations with reference to the debate about the nature of time in these
digital networks. Thereafter, we explain our experiment and its results. The
last section is devoted to the methodological and substantive conclusions.
Time
and the internet
In many different ways, the
internet has conveyed the notion that it somehow has a profound effect on the
relations between space and time. The early champions of the Net were convinced
of the breakdown of temporal and spatial differences by going online (Brand,
1987). Notions such as ‘timeless time’ (Castells, 1996, p. 464), ‘simultaneity
of non-simultaneous’ (Brose, 2004; Laguerre, 2004), ‘ultra-present’ (Goldhaber,
2004) and ‘extended present’ (Nowotny, 1994, p. 11) all aim at characterizing
the changes in our conceptions of time and temporality due to new ICTs and
digital networks.
Hassan (2003) has proposed the
notion of ‘network time’: Network time is digitally compressed clock-time, and
as such operates on a spectrum of technologically possible levels of
compression. This spectrum is ‘open ended’ (Hassan, 2003, p. 233). According to
Hassan, the observed acceleration of time follows from the premise of
asynchronicity among the networks, i.e., different frequencies of change: ‘The
“revolution” in information technologies has been to take this to another level
of temporality, to compress the meter of the clock and to accelerate the time
standard of modernity. The creation of the network has simultaneously created a
digital environment, an information ecology that generates its own temporality’
(Hassan, 2003, p. 233).
From this perspective the search
engines can be considered as subsystems of the e-society which function as
clocks of the internet that ‘tick’ at different frequencies. The search engines
update their catalogues at different frequencies, and as a consequence time is
reconstructed as a resonance effect between these different frequencies.
Whereas modern ‘clock time’ was designed to gather people at one place at the
same time, the internet would allow for simultaneous access to information free
from physical locations, thus leading to the ‘simultaneity of the
non-simultaneous’ (Brose, 2004; Laguerre, 2004). However, there are two
opposing views on how global networks affect the interplay between time and
space.
One side claims that global
networks lead to the dissolution of both time and space as relevant categories,
because everything can take place at the same time and largely independently of
geographical constraints. From this perspective, place is no longer relevant in
cyberspace. A more nuanced version of this position has been taken by Castells
(1996), who claimed that the measurable clock-time of the industrial revolution
is being shattered ‘in the network society, in a movement of extra-ordinary
historical significance.’ He captured this in the concept of ‘timeless time’
(Castells, 1996, p. 464): ‘I propose the idea that timeless time, as I
label the dominant temporality of our society, occurs when the characteristics
of a given context, namely, the informational paradigm and the network society,
induce systemic perturbation in the sequential order of phenomena performed in
that context.’ Brose (2004, pp. 16-17) argues that the impression of an
acceleration of time may be a result of the simultaneity of non-simultaneous,
multiple presents.
A second perspective claims that
the modernist clock-time, far from being dissolved, actually extends its
domination through ICT and the global networks. These scholars build on the
analysis of the role of technical time standardization in the rise of
capitalism and more specifically the industrial revolution (Thompson, 1967).
From this perspective, the central role of time has been the coordination (in
the sense of control and connecting) of social relationships (Elias, 1992). The
new digital technologies would play the same role, building on the social
process of standardization of time made possible by the mechanical clock (Adam,
2004). Urry (2000), for example, draws a parallel between the emergence of the
internet and the railway system in the 19th century.
Telegraphy first made it possible
to construct networks spanning the globe. Using international standard time
(GMT) these systems could be globalized (e.g., for the purpose of air traffic
control). These networks preceding the internet would already have extended the
domination of standard time to parts of the world that hitherto had been
relatively unaffected (Nowotny, 1994). Far from freeing individuals or groups
from the regime of the clock, the internet can be expected to subsume all
remaining variety to a new regime that is even stricter. This technical
standardization of time would leave no room for the post-modern deconstruction
of time (Adam, 2004).
In an exposé on the technicity of
time, Mackenzie (2001) proposed to conceptualize clock-time as a ‘temporal and
topological ordering that continues to unfold from a metastability.’ Mackenzie
compares time measurement to the sudden crystallization in a supersaturated
solution that makes the solution metastable. Metastability refers to the
tension in the synchronization of different ‘clocks’, and multiple presents. By
using this concept of metastability, Mackenzie (ibid.) wishes to combine three
analytical perspectives on time: Heidegger’s exteriorization of temporality,
Elias’s notion of the transitions between different social timing regimes, and
Latour’s view of the technical mediation of time. The two mechanisms of
processing in a forward mode and rewriting with hindsight can also be
distinguished in terms of the possibilities to stabilize or globalize a
metastability (Leydesdorff, 2001).
The dominance of linear time was
fueled by the industrial revolution, which enabled people to transform time
into money and place a premium on the rationalization of time.
Like the social construction of time, however, every conception of time should
take into account both its linear and cyclical dimensions. The present
re-conceptualization of time builds upon the standardized world time of the
industrial revolution, yet fundamentally alters it by adding cycles as older
notions of time. This reconceptualization is driven by the new information and
communication technologies as socio-technical practices. These technologies
generate a drive for ‘a world-wide condition of simultaneity’ (Nowotny, 1994,
p. 9). Because of the illusion that temporal and spatial differences matter
less, time and space seem to be compressed and collapsed in the world of the
internet into terms of globalized communications.
In summary, the concept of a single
time axis which is moving forward like an arrow is broken in the post-modern
appreciation of a variety of time horizons in different social systems and for
the different actors involved (Coveney & Highfield, 1990; Prigogine &
Stengers, 1988). Different updating and growth frequencies may resonate historically
into stability (e.g., institutions), and subsequently the metastability of the
resulting system can also be globalized into an order of expectations operating
in the present (Husserl, 1929; Luhmann, 2002). The present is not only the
fleeting, uncapturable moment between past and future, but also a broad horizon
of experiences in which pasts and futures are being recycled.
With an inspiration very similar to
that of Brose’s (2004) ‘simultaneity of the non-simultaneous,’ Goldhaber (2004)
describes the mentality of the Homo Interneticus as being captured in an
‘ultra-present’ where things constantly happen. The ultra-present is not only a
redefinition of the durée of the present, but also of a balance between
linearity and cyclicality. A comparable notion is captured by Nowotny’s concept
of the ‘extended present:’ ‘The permeability of the time-boundary between
present and future is increased by technologies which facilitate temporal
uncoupling and decentralization, and which produce different models of time referring
to the present that have largely become detached from linearity’ (Nowotny,
1994, p. 11). In short, the present can be considered as both the generator and
the result of interacting cycles that have their own specific frequencies. The
present of the search engines is created by the three updating frequencies of
the Web pages, the search engine databases, and the overall evolution of the
Web.
Perhaps, the internet can be seen
as the embodiment of an extended present, turned from really virtual to virtually
real thanks to the new technologies of virtualization (Latour, 1991). If this
were the case, we should add the notion of fragmentation to that of the
extended present because any resolution would necessarily remain historical. In
general, the reflexive operation contains a reference to the historical
situation, but that situation is looked at from the perspective of the present,
i.e., with hindsight. What is precisely added by the reflexive (albeit
automated) mechanism of rewriting the system (the internet) by a subsystem
(search engine) of the same system? Does the feedback arrow affect the
feedforward one, and if so, how? Perhaps we should amend the ‘extended present’
proposed by Nowotny (1994), and turn it into a notion of many competing and
fragmented, multiple extended presents—in the plural? The multiple extended
presents are a result of the resonances between the different updating cycles,
and this can be studied empirically by the analysis of search engine results.
We aim to study how this ‘present’ changes over time and across search engines.
Research
Design
Our experiments focus on how two
major search engines, AltaVista and Google, have reconstructed
the Web pages on ‘frankenfoods’ over time. The metaphor of ‘frankenfoods’ has
been used on the Web in the debate on genetically modified foods since the
mid-1990s in the pages of various consumer and environmental organizations, in
discussion forums and newsletters as well as in political arguments and
journalistic accounts of the debate. In these ‘static’, i.e. archived Web
pages, the use of the metaphor on the Web reached its
peak between 1998 and 2000, and thereafter its use decreased rapidly (Hellsten, 2003). In this study, we can
contrast this result with that of ‘dynamic’, i.e. faster changing Web pages as
represented in the search engine results. In other words, this search term
provides us with a well delineated topic and a relatively unambiguous search
term with a clear life cycle. It is interesting to see how the updating
mechanisms work on a topic on which new Web pages are not likely to have been
added since 2000, while the Web continues to grow all the time.
The data was initially collected on
21-23 January 2003 using only the AltaVista Advanced Search Engine. The
searches were at that time limited to the years 1995-2002. This data collection
was repeated exactly after one year, i.e., on 21-23 January 2004, and then
after one month, i.e., on 21-23 February 2004, and after three months, i.e., on
21-23 April 2004. The searches in 2004 used both AltaVista and Google,
and included the year 2003. The results for the year 2003 were further
decomposed into the twelve months of that year in order to distinguish between
the long-term and short-term effects of the updating in the different presents
in more detail.
The user interfaces of the two
search engines provide different options for using search terms. With AltaVista
we originally used the search string frankenfood* OR (frankenstein AND food*)
for the retrieval. We used the FaganFinder interface to Google
that allows us to use the date range capability of Google. However, this
interface does not allow the combination of Boolean operators, and the *
wildcard does not function in the ‘exact phrase’ option. For this reason, the
original search string was split into three versions, for each of which the
results were collected separately and then pooled: frankenstein food, frankenstein
foods and frankenfood(s). In order to
compare the results of Google with AltaVista, we also used the
following string in AltaVista: frankenstein food* OR frankenfood*
for the three searches conducted in 2004.
We not only checked the reported
number of hits of each search engine, but also downloaded the pages with the
search results. These pages contain the titles, first sentences, document
types, and URLs of the hits. This material allows us to check how many of the
reported results could actually be retrieved from the internet. More
importantly, the titles provide us with a semantic domain that can be mapped and
visualized in order to see how the words used in the titles of the results are
positioned, and whether the clusters of words change from one data collection
to another. We use techniques that were developed for this purpose in other
contexts (Leydesdorff, 2004: Leydesdorff & Hellsten, 2005) and provide the
visualizations below in order to illustrate our arguments with substantive
interpretations.
Our expectation about the changes
of the different presents generated by the search engines can be formulated as
follows. First, we expect that the distribution of the reported number of hits
over the years will show a strong bias in favour of the most recent year
(relative to the date of the measurement, i.e. the ‘present’ when the data was
collected). We call this the long-term memory of search engines. Second,
if it is true that Web sites are continuously overwritten with newer date
stamps, then we would expect a decrease in the total number of hits for
the months before the most recent one (again relative to the date of the
measurement). We call this the short-term memory.
In addition, we are interested in
the substance of the search results, i.e., the structure of the information
retrieved using the search engines. We tested if the structure in the data changes.
In summary, we study the construction of time both in terms of changes in the
reported numbers of results per year and the actually retrieved results. We use
the reported numbers for the study on short- and the long-term memory, while
the semantic maps are based on the retrieved results.
Re-Writing
the Past
Long-term memory
The long-term memory of the search
engines can be expected to show a bias towards the latest year. Figures 1 and 2
show the development of the long-term memory results of AltaVista and Google,
collected in January, February, and April 2004, respectively, and with exactly
the same strings (frankenfood* OR frankenstein food*).
Figures 1 and 2: Search
results using ‘frankenfood* OR frankenstein food*’ as search terms in Google
and AltaVista.
First, the Figures 1 and 2 show
that both AltaVista and Google have a strong and consistent bias
towards the latest year. For the measurement in January 2003 (including only AltaVista),
the year 2002 was the most recent year, hence this distribution is shifted one
year to the left (Figure 1). The data also prove that AltaVista and Google
both overwrite their histories. The sharpest fall is seen in the year 2002,
which can be attributed to a massive updating of Web pages from that year in
the year 2003, i.e., towards the closest to the present in question. The number
of hits retrieved using AltaVista decreased by an order of magnitude
after the search engine of Yahoo! was adopted. Google hits
decrease from the measurements in January to those in April 2004 for all years.
Except for the most recent year, the numbers in the measurements in February
and April are at the same level.
A decrease in the overall number of
results was expected because of our focus on the metaphor of ‘frankenfood.’ As
noted, the use of this metaphor reached a peak between the years 1998 and 2000
in ‘static’ Web pages, and no significant number of new Web pages was expected.
By using the search engines, one is able to detect also the changes in dynamic
Web pages. The search engine results are remarkable since they indicate that
even if the use of the metaphor has decreased, the Web pages are still updated
continuously. This is consistent with the results by Bar-Ilan (1999), who found
that as a result of this continuous updating, the search string is not always
necessarily a part of the Web page. The search engine
seems to ‘remember’ that a certain URL was part of its database index even if
the search term may have disappeared from the page.
Short-term memory
According to our hypothesis, the
search engines are expected to show a decrease in the total number of hits
towards the most recent months relative to the month of measurement, i.e., the
date of the data collection. Figures 3 and 4 show this short-term memory of the
search engines.
|
|
|
|
Figures 3 and 4: Search results using ‘frankenfood* OR Frankenstein food*’
as search terms in Google and AltaVista.
|
In the interpretation of these
graphs, one should take into account that our measurements were conducted in
the beginning of the year 2004 (January, February, and April). We would
therefore expect a bias in favour of the months December and November 2003 if
the search engines’ updating frequency is high enough to be noticeable within a
month. Both search engines do indeed record high numbers in the last months of
the previous year, but not with the same update frequency.
The Google data indicate a
shift over time towards the month of December. The AltaVista measurements
in January and February 2004 are at the same level, whereas the numbers in the
April measurement have fallen dramatically. We attribute this latter decline to
the take-over of AltaVista by Yahoo! The Google data,
however, substantiate the hypothesis that the historical record is being erased
in the short term: the numbers for the first ten months of the year
consistently decrease over time. AltaVista data does not show this
effect. Apparently, the two search engines differ in terms of the lags and
speeds in the updating of their databases.
Substantive similarities and differences
We also expected that the structure
of the Web pages would differ across the searches at different points of time.
When the documents are relocated to more recent years, one could expect that the
existing structure of documents within the year would be disrupted. Since this
relocation will not be uniform for all documents, a shift of the structure as a
whole into a more recent year is unlikely. Our expectation is therefore that
the structure, as represented by co-appearances in the title words of the
retrieved Web pages in a particular year, erodes over time. As a result, the
information in the structure of the results is gradually lost.
To study this, we first calculated
correlation coefficients, and then drew semantic maps based on the
co-occurrences of title words of the retrieved documents. The semantic maps are
based on asymmetrical matrices of word frequencies, where co-occurring words
are used as variables and the documents as cases. These matrices were imported
into UCINET and the visualizations were made with Pajek (for information on the
methods, see Leydesdorff, 2004; Leydesdorff & Hellsten, 2005).
Table 1 first summarizes the
numbers of all the downloads as contrasted to the reported numbers of results.
It also summarizes the numbers of the title words in the downloaded results and
the numbers of co-occurring title words, as well as the numbers of the
co-occurring title words included in the analysis.
|
|
AV Jan 2004
|
AV Feb 2004
|
AV Apr 2004
|
Ggl Jan 2004
|
Ggl Feb
2004
|
Ggl Apr
2004
|
|
Nr. of reported records
|
8222
|
8733
|
1239
|
5184
|
4955
|
3553
|
|
Nr of retrieved records
|
2106
|
2035
|
620
|
3068
|
2912
|
3115
|
|
% downloads
|
25.6
|
23.3
|
50.0
|
59.2
|
58.8
|
87.7
|
|
Unique title words
|
3495
|
3397
|
1821
|
4561
|
4328
|
4289
|
|
Nr. of word occurrences
|
9332
|
9004
|
3616
|
14597
|
14353
|
15020
|
|
Threshold used
|
> 12
|
> 12
|
> 6
|
> 18
|
> 18
|
> 18
|
|
Words included in the
semantic maps
|
97
|
95
|
74
|
98
|
103
|
111
|
|
Cosine ≥ 0.2
|
46
|
49
|
57
|
47
|
57
|
67
|
Table 1 Summary of the downloads
Note the discrepancies between the
reported numbers and the actually downloaded numbers in the first two rows of
Table 1. These differences are the more striking since we used the ‘site
collapse’ filter of the AltaVista Advanced Search Engine in order to
exclude identical pages from the reported numbers. AltaVista provided
higher reported numbers in January and February than Google, yet Google
had higher numbers of actually downloadable records. It is intriguing that the
actually downloaded numbers decrease much more slowly than the reported
numbers. In the case of Google in April 2004 this number even rises,
resulting in an exceptionally high percentage of downloaded records. The number
of retrievable records seems to remain approximately the same, but the reported
number may become more precise over time.
The rank order in the frequencies
of words in the titles of Web pages are an indication of the similarity of
structure in the Web search on ‘frankenfoods’ time-stamped for the year 2003.
This information is summarized as rank-order correlations in Table 2.
|
|
AV Jan
|
AV Feb
|
AV Apr
|
Ggl Jan
|
Ggl Feb
|
Ggl Apr4
|
|
AV Jan
|
|
1.000
|
.840(**)
|
.460(**)
|
.525(**)
|
.524(**)
|
.530(**)
|
|
|
N
|
3351
|
2965
|
959
|
1918
|
1806
|
1720
|
|
AV Feb
|
|
.840(**)
|
1.000
|
.495(**)
|
.564(**)
|
.556(**)
|
.541(**)
|
|
|
N
|
2965
|
3247
|
958
|
1902
|
1775
|
1689
|
|
AV Apr
|
|
.460(**)
|
.495(**)
|
1.000
|
.531(**)
|
.533(**)
|
.519(**)
|
|
|
N
|
959
|
958
|
1766
|
1168
|
1195
|
1172
|
|
Ggl Jan
|
|
.525(**)
|
.564(**)
|
.531(**)
|
1.000
|
.696(**)
|
.623(**)
|
|
|
N
|
1918
|
1902
|
1168
|
4360
|
2978
|
2537
|
|
Ggl Feb
|
|
.524(**)
|
.556(**)
|
.533(**)
|
.696(**)
|
1.000
|
.815(**)
|
|
|
N
|
1806
|
1775
|
1195
|
2978
|
4145
|
3448
|
|
Ggl Apr
|
|
.530(**)
|
.541(**)
|
.519(**)
|
.623(**)
|
.815(**)
|
1.000
|
|
|
N
|
1720
|
1689
|
1172
|
2537
|
3448
|
4077
|
** Correlation is significant at the 0.01 level (2-tailed).
Table 2: Rank correlation (Spearman’s ρ) among word occurrences in
different searches for the documents from the year 2003.
The word frequency distributions
are most strongly correlated within each search engine, with the exception of AltaVista
in April 2004. This auto-correlation reflects that the search engines are more
likely to update URLs already included in their databases than to add new
entries. Furthermore, the overlap between the search engines is expected to be
moderate because of the different algorithms used for the crawling. The
correlation coefficients prove that the change in AltaVista in April
2004 has fundamentally changed the operation of this search engine. Because of
this structural change, we will only show the semantic maps for AltaVista
as collected in January and February 2004. The correlations between AltaVista
and Google are lower than those between the structures generated by each
search engine, but their significance cannot be ignored.
What clusters can be discerned in
these two related structures? The next two figures show the structure of the
title words that co-occurred more than twelve times in the search results
retrieved using AltaVista. The semantic maps are based on the searches
conducted with AltaVista in January and February 2004 and represent the
results attributed by the date stamps to the year 2003 (Figures 5 and 6).
Figure 5: 46 words related at the level of cosine ≥ 0.2 and occurring more
than 12 times in the 2106 records collected with the AltaVista Advance
Search Engine in January 2004.
Figure 6: 49 words related at the level of cosine ≥ 0.2 and occurring more
than 12 times in the 2035 records collected with the AltaVista Advance
Search Engine in February 2004.
The picture of January is more
informative in the sense of representing more connected—that is,
larger—clusters than the one from February. Thus, the structure of the
information erodes in the data over time. Fewer words explain more structure in
data collected in January than in February. For example, the number of
unrelated clusters increases from eight in January to twelve in February. At
the same time, the results indicate continuity in the data. For example, the
word clusters around Earth Press Release and Plant Forum Archive in both
January and February are similar. Is this erosion also visible in the
structures generated by Google (Figures 7 to 9)? The structure of the
co-occurring title words in the set retrieved by Google clearly differs
from that of AltaVista by representing unique word clusters around
‘public debate’ and ‘new global headlines.’ However, some clusters reflect the
same words, such as Plant Forum Archive (Figure 7). The Web pages in this
archive are stable.
Figure 7: 47 words related at the level of cosine ≥ 0.2 and occurring more
than 18 times in the 3068 records collected with the Google Search Engine
in January 2004.
Figure 8: 57 words related at the level of cosine ≥ 0.2 and occurring more
than 18 times in the 2912 records collected with the Google Search Engine
in February 2004.
Between January and February, Google
shows a change in terms of this analysis more than AltaVista did. Figure
8 exhibits the emergence of new word clusters, such as ‘characters that cannot
be correctly displayed,’ ‘time online,’ ‘Web blog,’ and ‘group message’ that
were not part of the title words collected in January. How the structure of the
results further develops is presented in Figure 9.
Figure 9: 67 words related at the level of cosine ≥ 0.2 and
occurring more than 18 times in the 3115 records collected with the Google
Search Engine in April 2004.
Figure 9 shows the further erosion
of the structure of the co-occurring title words, and the emergence of one new
cluster around ‘education portal.’ The number of disconnected clusters in the
map rises from ten in January to fourteen in April.
In summary, the picture is a bit
more complex than in the case of AltaVista, mainly because of the
emergence of a cluster of words in February that is related to the message
‘contains characters that cannot be correctly displayed.’ Nevertheless, the
trend for both AltaVista and Google is that of a loss of
structure.
The maps based on the Google
data appear more ‘noisy’ than those made from the January and February AltaVista
data. This may be caused by the fact that the search engine of Google is
based on using hyperlinks, whereas AltaVista uses keyword searches. A
second difference between Google and AltaVista is shown in the
cluster about non-displayable (i.e., non-latin) characters in the titles of the
results. The algorithm of the search engine may have changed. We tentatively
infer from this that Google has a wider window on non-English languages,
precisely because it is not based on (English) keyword searches. Another
explanation for this might be that Google indexes a wider variety of
file types than most other search engines, for example, including pictures in
the database (Reddy & Wouters, 2003). Furthermore, one should keep in mind
that the search engines may crawl other parts of theWeb
Discussion
We had three hypotheses about how
the present of the search engines evolves over time. First, we expected that
the distribution of the reported number of hits over the years 1995-2003 would
show a strong bias in favour of the most recent year (relative to the date of
the data downloading). Secondly, if it is true that Web sites are constantly
overwritten with newer date stamps, then we expected also to find a decrease in
the total number of hits for the months before the most recent one (again
relative to the date of the measurement). The results confirm both these
hypotheses. We expected substantial erosion over time in the sense that search
engines not only re-write the past but also forget the past. This dynamics of
reconstruction from the perspective of hindsight is extremely relevant because
the reorganization feeds back on the overall growth of the Web. Thirdly, we
expected that the structure of the information erodes over time, and our results
confirmed this. The past in the internet is constantly overwritten from a
hindsight perspective that affects the numbers of the results as well as the
actual Web pages the search engines retrieve. Hence, the presents from where
the data is collected affect the search results considerably.
Our third finding—that both search
engines not only lose information quantitatively, but that they also erase the
structures entailed in the relationships between words of Web page titles—may
be even more important from a social science perspective than the updating of
the time stamps of Web pages and Web documents as such. What is particularly
striking is that we are not dealing here primarily with instabilities. On the
contrary, in many ways the updating mechanism of search engines is remarkably
stable and systematic.
This does not mean we did not meet
any instabilities. We experienced two types of instabilities: first, the
fundamental restructuring of AltaVista in April 2004, which made its
results before and after the re-organisation unrelated; and second, the
variable difference between the reported number of records and the number of
records that could actually be retrieved.
However, the main phenomenon we
have dealt with in our experiments is not instability, but the systematic
erasure of both the historical record and the structures in informational and
semantic networks. This is caused by the fact that the search engines are tied
to the updating cycles of the Web and the internet, rather than to the historical
development of its structure. The structure in the information on the Web at
any given moment of time is the result of relations that exist at the moment of
the creation of Web pages and the various updating mechanisms that we have
shown in our experiments. Although we do not know how these two forces balance
each other during the period of incorporating the Web pages into the search
engines, the lag times seem different between the two search engines under
study. However, we have shown that over time, the structure of information as
represented in the relationships between words is determined by the updating
frequencies, and as a consequence, erosion of this structure is taking place.
In other words, the fact that the
search engines of the internet and the Web are actually a complex network of
networks, each with its own updating cycles, leads to a loss of structure in
the word clusters. This raises the question of whether this loss of structure
may have a finite window: are networks of information after a longer period of
time becoming more stable, or does the erosion of structure continue? We would
expect the latter to prevail because of the continuation of the relevant
operations. Another relevant question would be whether particular configurations
of networks ‘travel’ from year to year. It would also be interesting to know to
what extent the specifics of the search strings (‘frankenfoods’ in our case)
influence the types of networks generated and their decay times.
Conclusion
We have shown that the search
engines AltaVista and Google systematically relocate the time
stamps of Web documents in their databases from the more distant past into the
present and the very recent past. Second, the search engines delete documents
from the year to which they were initially assigned. This leads to a loss of
information in the historical record on the Web as represented in the search
engine databases. Third, information is lost not only in the quantitative sense
of documents disappearing from the historical record, but also in the sense of
a loss of structure in the semantic networks.
This does not mean that search
engines are completely useless for scholarly research or do not represent a
significant phenomenon on the Web. On the contrary, our results confirm that we
can appreciate search engines as the clocks of the internet, ticking at
different frequencies and possibly leading to multiple presents. They provide
representations that indicate the updating frequencies of both the Web and the
underlying internet. How should we interpret this? We are dealing with complex
interactions of the updating frequencies of individual Web pages by their
authors or Web masters; the updating frequencies and mechanisms of the
structures in which these Web pages are positioned; the frequencies with which
these Web pages are being visited by search engine crawlers; the extent to
which ‘old’ Web pages are retained in search engine databases although more
recent version of that ‘same’ Web page have been added; the overall growth of
the Web; and, last but not least, by the updating frequencies of the sorting
algorithm of the search engine and its presentation mechanisms. All these
frequencies can be expected to differ. Moreover, each search will be influenced
in different ways by the various frequencies. Each search engine can therefore
be said to represent not one updating frequency but a frequency distribution or
spectrum of frequencies (including very slow changes for static Web pages).
This spectrum may be specific to a given search engine in a particular period
of its existence (Smolensky, 1986).
What does this mean in relation to
time and temporality? As clocks of the internet, search engines realize the
present as a collection of extended presents that can exist in parallel on the
Web. In other words, time is being represented as realities that coexist in
space. The concept of the ‘extended present’ (Nowotny, 1994) has been used
mainly to indicate the dominance of the present over the past and future, and
to broaden the concept of the present from a fleeting point in time to a
spectrum of actualities. We propose to add the notion of fragmentation to this
concept. We are not dealing with one extended present, but with a multiplicity
of partly conflicting presents. The frequency with which these extended
presents are being updated in turn does not have to relate to the development
of the actors whose presence is here represented. This differentiation may
contribute to the feeling of being overwhelmed by the information overload of
the system (Luhmann, 1996).
We have interpreted search engines
as clocks of the internet driven by the interaction of a variety of updating
frequencies. We have shown with our experiments that these clocks not only run
at different frequencies depending on the ‘present’ of the searches and the
search engine in question, but also reconstruct the pasts in very different
terms. Each search engine differs in the combination of these frequencies and
their selection, resulting in different lag times and information restructuring
windows. The question of how we can make better use of search engines in
scholarly research to unveil the overall updating cycles that dominate the Web
and particular domains of the Web becomes an interesting research question that
should be put on the agenda. Search engines are the ‘clocks’ of the Web, but
rather strange ones that act more like the clocks of Salvador Dali than those
of Christiaan Huygens.
return
References
Adam,
B. (2004) Time. Cambridge: Polity Press.
Bar-Ilan,
J. (1999) ‘Search Engine Results over Time - A Case Study on Search Engine
Stability.’ Cybermetrics, 2/3, URL (consulted Nov. 2004) http://www.cindoc.csic.es/cybermetrics/articles/v2i1p1.html.
Bar-Ilan,
J. (2001) ‘Data collection methods on the Web for informetric purposes - A
review and analysis.’ Scientometrics, 50(1): 7-32.
Bar-Ilan,
J., and B. C. Peritz (2002) ‘Informetric Theories and Methods for Exploring the
Internet: An Analytical Survey of Recent Research Literature.’ Library
Trends, 50(3): 371-392.
Brand,
S. (1987) The media-lab. Inventing the future at M.I.T. New York: Penguine
Books.
Brewington,
B. and Cybenko, G. (2000) ‘Keeping up with the changing web.’ Proceedings of
the 9th International World Wide Web Conference, pp. 257-276.
Brose,
H.-G. (2004) ‘An introduction towards a culture of non-simultaneity?’ Information
Society 13(1): 5-26.
Castells,
M. (1996) The rise of the network society. Cambridge, MA: Blackwell.
Coveney
P. & Highfield R. (1990) The Arrow of Time. W.H.Allen.
Dubois,
D. M. (1998) Computing Anticipatory Systems with Incursion and
Hyperincursion. In D. M. Dubois (Ed.), Computing Anticipatory Systems, Casys-First
International Conference (Vol. 437, pp. 3-29). Woodbury, NY: American
Institute of Physics.
Elias,
N. (1992) Time: An Essay. Oxford: Blackwell .
Goldhaber,
M. (2004) The mentality of Homo Interneticus. First Monday 9(6).
Hassan,
R. (2003) ‘Network Time and the New Knowledge Epoch.’ Time & Society, 12(2/3):
225-241.
Hellsten,
I. (2002) The Politics of Metaphor: Biotechnology and
Biodiversity in the Media. Acta
Universitatis Tamperensis; 876, Tampere University Press, Tampere. Partly
available on-line at: http://acta.uta.fi/pdf/951-44-5380-8.pdf
Hellsten,
I. (2003) ‘Focus on Metaphors: The case of “Frankenfood” on the web.’ Journal
of Computer-Mediated Communication 8(4) July, URL (consulted Nov. (2004): http://www.ascusc.org/jcmc/vol8/issue4/hellsten.html.
Husserl, E. (1929) Cartesianische Meditationen Und Pariser Vorträge.
[Cartesian Meditations and the Paris Lectures.]. The Hague: Martinus Nijhoff, 1973.
Innis, H. A. (1952) Changing Concepts of Time. Oxford, UK: Rowman & Littlefield, 2004.
Introna,
L. D. and H. Nissenbaum (2000) ‘Shaping the Web: Why the Politics of Search
Engines Matters.’ The Information Society, 16: 169-185.
Laguerre,
M. S. (2004) ‘Virtual Time - The processuality of the cyberweek.’ Information,
Communication & Society, 7(2): 223-247.
Latour,
B. (1988) Science in Action. How to follow scientists and engineers thorugh
society. Harvard UP.
Latour, B. (1991) Nous n’avons jamais été modernes. Paris: La
Décourverte.
Leydesdorff,
L. (1994) ‘Uncertainty and the Communication of Time.’ Systems Research,
11(4): 31-51.
Leydesdorff,
L. (2001) A Sociological Theory of Communication: The Self-Organization of
the Knowledge-Based Society. Parkland, FL: Universal Publishers, URL
(consulted Nov. 2004) http://www.upublish.com/books/leydesdorff.htm
Leydesdorff,
L. (2004) ‘The University-Industry Knowlege Relationship: Analyzing Patents and
the Science Base of Technologies.’ Journal of the American Society for
Information Science & Technology, 55(11): 991-1001.
Leydesdorff,
L. (2005) Anticipatory Systems and the Processing of Meaning:
A Simulation
Inspired by Luhmann's Theory of Social Systems. Journal of
Artificial
Societies and Social Simulation, 8(2), Paper 7, at
http://jasss.soc.surrey.ac.uk/8/2/7.html.
Leydesdorff,
L. & Hellsten, I. (2005) ‘Metaphors and Diaphors in Science Communication.
Mapping the Case of Stem cell Research.’ Science Communication 27(1):
1-36.
Lewandowski,
D. (2004) ‘Date restricted queries in web search engines’. Online
Information Review 28(6).
Luhmann, N. (1996) Die Realität der Massenmedien. Opladen:
Westdeutscher Verlag.
Luhmann,
N. (2002) ‘The Modern Sciences and Phenomenology,’ in W. Rasch (ed.), Theories
of Distinction: Redescribing the Descriptions of Modernity, pp. 33-60. Stanford, CA: Stanford University Press.
Mackenzie,
A. (2001) ‘The Technicity of Time.’ Time & Society, 10(2/3): 235-257.
Nowotny,
H. (1994) Time. Cambridge: Polity Press.
Prigogine, I., & I. Stengers. (1988) Entre le temps et
l'éternité. Paris: Fayard.
Reddy, C. and P. Wouters (2003) The invisible web in the EU
research area. European Comission, WISER project, deliverable 5-1, August 2003
Risvik,
K & Michelsen, R. (2002) ‘Search engines and Web dynamics’. Computer
Networks 39: 289-302.
Rosen,
R. (1985) Anticipatory Systems: Philosophical, Mathematical and
Methodological Foundations. Oxford, etc.: Pergamon Press.
Salton,
G., and M. J. McGill. (1983) Introduction to Modern Information Retrieval.
Auckland, etc.: McGraw-Hill.
Schütz, A. (1932) Der Sinnhafte Aufbau der Sozialen Welt (G.
Walsh & F. Lehnert, Trans. Vol. 1932). Vienna: Springer.
Smolensky,
P. (1986) ‘Information Processing in Dynamical Systems: Foundation of Harmony
Theory,’ in D. E. Rumelhart, J. L. McClelland & the PDP Research Group
(eds.), Parallel Distributed Processing, Vol. I, pp. 194-281. Cambridge, MA/ London: MIT Press.
Thelwall, M. (2001) The Responsiveness of Search Engine Indexes. Cybermetrics, 5(1), http://www.cindoc.csic.es/cybermetrics/articles/v5i1p1.html.
Thelwall,
M. (2002) ‘Methodologies for Crawler Based Web Surveys.’ Internet Research:
Electronic Networking and Applications, 12(2): 124-138.
Thompson,
E. P. (1967) ‘Time, Work-Discipline, and Industrial Capitalism.’ Past and
Present, 38: 56-97.
Urry,
J. (2000) Sociology Beyond Societies :
Mobilities for the Twenty First Century
Van Couvering, E. (2004) ‘New media? The political economy of
Internet Search Engines,’ paper presented at the AoIR conference, Brighton, UK, 19-22 September.
Vaughan,
L. and M. Thelwall (2004) ‘Search engine coverage bias: evidence and possible
causes,’ Information Processing & Management, 40(4): 693-707.
Wouters,
P., and D. Gerbec (2003) ‘Interactive Internet? Studying Mediated Interaction
With Publicly Available Search Engines.’ Journal of Computer Mediated
Communication, 8(4), URL (consulted Nov. 2004): http://www.ascusc.org/jcmc/vol8/issue4/wouters.html
Wouters,
P., I. Hellsten, and L. Leydesdorff. (2004). ‘Internet Time and the Reliability
of Search Engines.’ First Monday, 9(10), URL (consulted Nov. 2004): http://www.firstmonday.org/issues/issue9_10/wouters/index.html.
return
<pdf>
